Humans are one step closer to traveling at faster-than-light speeds. In a surprising paper, scientists say they've nailed down a physical model for a warp drive, which flies in the face of what we've long thought about the crazy concept of warp speed travel: that it requires exotic, negative forces. To best understand what the breakthrough means, you'll need a quick crash course on the far-out idea of traveling through folded space. The colloquial term “warp drive” comes from science fiction, most famously Star Trek. Star Trek suggests that this extraordinary power alone pushes the ship at faster-than-light speeds. Scientists have been studying and theorizing about faster-than-light space travel for decades. One major reason for our interest is pure pragmatism: without warp drive, we're probably never making it to a neighboring star system. The closest such trip is still four years long at light speed. Our current understanding of warp speed dates back to 1994, when a now-iconic theoretical physicist named Miguel Alcubierre first proposed what we've called the Alcubierre drive ever since. “By a purely local expansion of spacetime behind the spaceship and an opposite contraction in front of it,” Alcubierre wrote in his paper's abstract, “motion faster than the speed of light as seen by observers outside the disturbed region is possible.” Essentially, an Alcubierre drive would expend a tremendous amount of energy—likely more than what's available within the universe—to contract and twist space-time in front of it and create a bubble. Alcubierre describes spacetime expanding on one side of the ship and contracting on the other, thanks to that enormous amount of energy and a requisite amount of exotic matter—in this case, negative energy. Some scientists have criticized the Alcubierre drive, however, because it requires too much mass and negative energy for humans to ever seriously construct a warp-based propulsion system. NASA has been trying to build a physical warp drive through Eagleworks Laboratories for most of the last decade, but hasn't yet made any significant strides. Plus, Alcubierre himself has endorsed the new model, which is like having Albert Einstein show up to your introductory physics class. Here's a helpful video in which Sabine Hossenfelder, a Professor and Research Fellow at the Frankfurt Institute for Advanced Studies, breaks down the findings: Of course, there's one gigantic caveat here: The concept in this paper is still in the “far future” zone of possibility, made of ideas that scientists still don't know how to construct in any sense. “While the mass requirements needed for such modifications are still enormous at present,” the APL scientists write, “our work suggests a method of constructing such objects based on fully understood laws of physics.” Caroline Delbert is a writer, avid reader, and contributing editor at Pop Mech. Why Everything We Know About Gravity May Be Wrong The Source of All Consciousness May Be Black Holes Why Time Reflections Are a ‘Holy Grail' in Physics

The two-masted schooner was fully intact—a rarity in Great Lakes shipwreck finds. A team of five divers descended over 300 feet into Lake Ontario's dark waters near Toronto and found much more than they were dreaming of. The divers discovered a two-masted schooner sitting upright, with both masts still fully intact, in what the Ontario Underwater Council called an “extraordinary state of preservation for a Great Lakes vessel.” After some preliminary investigation, the team believes the ship may come from an under-documented period of Great Lakes shipbuilding, sometime between 1800 and 1850. “It took us a few moments to calm ourselves down because it's overwhelming finding a pristine wreck that is all in one piece,” Heison Chak, the president of the Ontario Underwater Council, told the CBC. We saw two—both masts were standing, which is pretty rare. A fiber-optic cable survey of the lake from Buffalo to Toronto alerted experts to an anomaly sitting on the lakebed, leading experts to surmise it could be the Rapid City vessel, an 1884-built schooner lost in 1917. Now, though, they don't think it can be that recent a wreckage. James Conolly, Trent University archaeologist and diver, said there were features that just weren't common for ships built after 1850, a period that experienced a bit of a technological leap for Great Lakes ships. Post-1850s ships had metal rigging, whereas the one found is rope-rigged. “This is deep enough that I don't think anyone's been on it,” Chak said. “I think we're the first group and that joy was just overwhelming.” Charles Beeker, a Great Lakes shipwreck expert and professor at Indiana University, told the CBC it's too early to say the vessel is from 1800 to 1850. “I don't want to diminish the value of it,” he said. “Where a wreck might once have survived intact for centuries,” Conolly said, “we now have only decades to study it before biological and environmental factors take their toll.” “We don't know, but if it is really that era,” Chak said, “from 1800 to 1850, I think we will have even more celebration because we hit the jackpot where there is very little history or studied documented material about shipwrecks, ships, or shipbuilding in that era.” Tim Newcomb is a journalist based in the Pacific Northwest. He covers stadiums, sneakers, gear, infrastructure, and more for a variety of publications, including Popular Mechanics.

You are using a browser version with limited support for CSS. On a video call in early September, Sarah Tabrizi first saw the data that she and other researchers studying Huntington's disease had been chasing for decades: compelling evidence that a gene-targeting therapy could slow the relentless progression of the neurodegenerative brain disorder. Before these results, “I was beginning to get a little bit worried that maybe, by the time people develop symptoms, that it was going to be too late to treat”, says Tabrizi, a neurologist who directs the Huntington's Disease Centre at University College London. “It's a giant step forward,” says Tabrizi, who was the trial's lead scientific adviser. The first-in-class gene therapy — called AMT-130 and developed by uniQure, a biotechnology company in Amsterdam — uses a harmless virus to deliver strands of genetic material into affected brain regions. On a standard rating scale used to assesses motor and cognitive functions and other measures of daily living, the scores of participants receiving the high dose dropped by just 0.38 points over three years. After digesting the findings, Tabrizi and her close collaborator Ed Wild, a fellow neurologist at University College London, shared what Wild describes as a “massive hug”. But, after that, it was right back to the daily demands of patient care and research. “We're like an episode of The West Wing: it's always, ‘What's next? Next for Tabrizi and Wild are leading roles in evaluating five other huntingtin-lowering therapies in clinical development, along with several others poised to enter human trials soon. “Sarah is amazing,” says Hugh Rickards, a neuropsychiatrist at the University of Birmingham, UK. You name a disease-modifying therapy in HD — she's got her hand on it somewhere.” Plus, “she's one of the nicest people you'll ever meet”, says Samuel Frank, a neurologist at the Beth Israel Deaconess Medical Center in Boston, Massachusetts, who, like Rickards, has worked with Tabrizi on Huntington's trials. Only four years ago, another promising huntingtin-targeted therapy, tominersen, faltered in late-stage trials. Overall, the drug failed to improve people's outcomes compared with those of the control group and it came with dangerous side effects at higher doses. Huntington's disease treated for the first time using gene therapy Stop the nonsense: genome editing creates potentially therapeutic transfer RNAs Men's brains shrink faster than women's: what that means for Alzheimer's The Dean of Systems Hub holds both an executive and faculty appointment and serves as a key member of the University's senior administrative team. CMLR's goal is to advance machine learning-related research across a wide range of disciplines. Build your own independent research group with long-term support at the Xuemin Institute, Fudan University! An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Hawaii's Kilauea Volcano Absolutely Destroys This Webcam in a Fiery New Video Hawaii's Kilauea, one of Earth's most active volcanoes, sent lava fountains spewing into the air, obliterating a U.S. Geological Survey camera Lava fountains are seen at both the north and south vents of Kilauea's summit around 10:00 A.M. local time on December 6, 2025. If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. The volcano is going through a long, intermittent eruption that began on December 23, 2024. The lava poured through residential neighborhoods, destroying hundreds of homes and other buildings. Andrea Thompson is senior desk editor for life science at Scientific American, covering the environment, energy and earth sciences. She has been covering these issues for nearly two decades. Prior to joining Scientific American, she was a senior writer covering climate science at Climate Central and a reporter and editor at Live Science, where she primarily covered earth science and the environment. She has moderated panels, including as part of the United Nations Sustainable Development Media Zone, and appeared in radio and television interviews on major networks. in atmospheric chemistry from the Georgia Institute of Technology. If you enjoyed this article, I'd like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history. I hope it does that for you, too. If you subscribe to Scientific American, you help ensure that our coverage is centered on meaningful research and discovery; that we have the resources to report on the decisions that threaten labs across the U.S.; and that we support both budding and working scientists at a time when the value of science itself too often goes unrecognized. In return, you get essential news, captivating podcasts, brilliant infographics, can't-miss newsletters, must-watch videos, challenging games, and the science world's best writing and reporting. There has never been a more important time for us to stand up and show why science matters. I hope you'll support us in that mission.

A Vitamin Shot Given at Birth Prevents Lethal Brain Bleeds, but More Parents Are Opting Out Vitamin K injections have prevented deadly brain bleeds in infants for more than 60 years. New research shows refusal rates have recently jumped nearly 80 percent The cause: vitamin K deficiency—a condition that is almost entirely preventable with a simple shot given at birth. According to new research published today in JAMA, the rate of vitamin K shot refusals has risen nearly 80 percent in the U.S. between 2017 and 2024. The trend comes at a time of growing vaccine hesitancy and science denial. Wellness influencers have fueled skepticism by characterizing the vitamin K shot as unnecessary and questioning its lab-made ingredients, even though they are found in many routine injections. If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. Without the shot, as many as one in 60 babies are at risk of vitamin K deficiency bleeding, which can cause permanent brain damage or death. “We know unequivocally that infants that don't receive vitamin K are at significantly higher risk of getting serious bleeding,” Scott says. His study did not directly measure whether the increase in refusals has led to more bleeding events. Some online influencers suggest that parents can use oral vitamin K as an alternative to the injection. But Scott says that approach is less reliable—oral vitamin K is less effective at preventing bleeds because it's not absorbed as readily. The shot itself carries minimal risk, and adverse reactions are rare. Loyal, who works at Yale New Haven Children's Hospital and has also researched vitamin K refusal, emphasizes that the shot differs fundamentally from a vaccine because it doesn't target a microorganism—it simply supplies an essential vitamin. Although the risk of vitamin K deficiency bleeding may seem low, Loyal says, “it's not zero—and why take a chance when we know there is a safe preventative measure?” Christina Szalinski is a freelance science writer who covers life sciences and health. If you enjoyed this article, I'd like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history. If you subscribe to Scientific American, you help ensure that our coverage is centered on meaningful research and discovery; that we have the resources to report on the decisions that threaten labs across the U.S.; and that we support both budding and working scientists at a time when the value of science itself too often goes unrecognized. In return, you get essential news, captivating podcasts, brilliant infographics, can't-miss newsletters, must-watch videos, challenging games, and the science world's best writing and reporting. There has never been a more important time for us to stand up and show why science matters.

While horoscopes and New Age crystals often provide clear answers to the first two questions, the last one on that list can get complicated sometimes. Type O positive uses a different antigen altogether (known as the rhesus, or Rh, factor), and Type O negative has none of the above. All of this complexity is encoded into proteins by out ABO gene, located on our ninth chromosome. While interesting in and of themselves, antigens become especially important when it comes to donating blood to and receiving blood from others. This is whytype O negative blood is seen as the universal donor type—because it has neither of the A or B antigens nor the Rh(D) antigen, there is virtually no risk of rejection. But even rarer blood types than O negative exist. There are fifty people on the entire planet with Rh-negative blood, nicknamed “golden blood,” which is completely devoid of Rh antigens. And there is only one individual known to have Gwada negative blood, which was only discovered earlier this year. While O negative is only found in 7% of the global population, researchers who tested the blood of patients and donors in a Thai hospital have now discovered yet another new type that runs through the veins of only three known individuals (it was found in one patient and two donors). “ABO discrepancies were distinct between donors and patients even in the same ethnicity,” the researchers said in a study recently published in the journal Transfusion and Apheresis Science. Additionally, the B(A) individuals identified in this study carried identical genetic alterations that differed from all antecedent alleles of the B(A) phenotype.” The researchers saw four alleles—alternative forms of certain genes at the same location on a chromosome—on the B(A) type that differed from alleles associated with other blood type and created B(A). Mutations like this had previously been seen in many ethnicities, including individuals from other parts of Asia, but it was the first time such a phenomenon was ever recorded in the Thai population. There could be more blood types out there that remain undiscovered, and more individuals with those extremely rare blood types. Further research may find more of either, and this discovery is proof that blood typing is not so simple as ABO. Her work has appeared in Popular Mechanics, Ars Technica, SYFY WIRE, Space.com, Live Science, Den of Geek, Forbidden Futures and Collective Tales. She lurks right outside New York City with her parrot, Lestat. When not writing, she can be found drawing, playing the piano or shapeshifting. Experts May Have Found Why We Gained Consciousness

Included in the haul is a cannon, two porcelain cups, and three gold and bronze coins. But there's plenty more still sitting 1,970 feet below the ocean's surface off the coast of Colombia, as much as $17 billion worth. The ship was laden with 10 years' worth of gold, silver, and gems, tribute from Spanish colonies in Latin America headed back to the Spanish king. Experts have now pulled up the first pieces from the ship, including coins, a cannon, and porcelain cups. where it laid dormant for hundreds of years. An estimated 200 tons of gold, silver, and uncut gemstones were aboard the ship, the result of 10 years of taxation saved up before the fleet's planed voyage back to Spain. These coin hoards likely formed part of the royal treasure dispatched from Peru. But when the fleet of 18 ships left Cartagena bound for Spain on June 8, 1708, it was attacked by five British warships. Tim Newcomb is a journalist based in the Pacific Northwest. He covers stadiums, sneakers, gear, infrastructure, and more for a variety of publications, including Popular Mechanics. Archaeologists Found a Medieval Knight's Odd Skull This Decade-Old Whale Carcass Still Supports Life A Whole New Cat Color Has Emerged Experts Found 100 Structures From an Ancient City This Ancient Tomb Secretly Held Famous Royals Odd Signals in Antarctica Are Baffling Scientists Experts May Have Found Why We Gained Consciousness

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. (2025)Cite this article Organometallic reagents are essential in organic synthesis, with organolithium compounds being most widely used. However, as lithium becomes less abundant and increasingly expensive, organosodium compounds have emerged as promising alternatives, but their use in organic synthesis is limited by their poor solubility in organic solvents, the need for pre-activated sodium sources and the necessity for highly anhydrous conditions. Here we report a mechanochemical protocol for the direct generation of organosodium compounds from cheap and shelf-stable sodium lumps and readily available organic halides under bulk, solvent-free conditions. These reactions generate an array of organosodium compounds in minutes, without special precautions against moisture or temperature control. These nucleophiles can be used directly for one-pot nucleophilic addition reactions with electrophiles and nickel-catalysed cross-coupling reactions. Furthermore, this mechanochemical approach enables the sodiation of inert C–F bonds in organic fluorides. This method is anticipated to drive progress in sodium-based synthetic chemistry. Organolithium compounds have played a dominant role as carbon nucleophiles, Brønsted bases and precursors for a variety of other organometallic reagents in organic synthesis for over 100 years (Fig. Their widespread use is attributed to their relatively high stability in organic solvents, coupled with well-established methods for their preparation5,6. However, from a sustainability standpoint, there is an increasing demand for alternatives to lithium, which is becoming both less abundant and more expensive7. The growing demand for lithium-ion batteries is expected to further intensify competition for lithium resources8,9,10. a, Comparison between organolithium compounds and organosodium compounds. b, Previous work involving mechanochemical activation of sodium metal. c, Direct mechanochemical synthesis of organosodium compounds through ball milling. E+, electrophile; SCXRD, single-crystal X-ray diffraction. Sodium, another s-block metal, is much more earth abundant than lithium (crustal abundance 22,700 ppm versus Li abundance 18 ppm)11 and is therefore an attractive candidate for replacing lithium to explore sustainable organic synthesis (Fig. The first reports of organosodium chemistry emerged in the 1850s, pioneered by J. A. Wanklyn24, but substantial challenges exist that have prevented the widespread adoption of organosodium reagents within the field of organic synthesis25,26,27,28,29. First, the highly reactive nature of organosodium reagents makes them incompatible with many solvents, including ethers, which can react via C‒H and C‒O activation, and even aromatic solvents such as toluene and benzene29. Only alkane solvents (for example hexanes) can be considered genuinely inert towards organosodiums, but the poor solubility of organosodium reagents in hydrocarbon solvents often causes substantial problems, limiting the potential for widespread application. Second, while direct metallation of organic halides with metallic sodium is an attractive and straightforward synthetic method, it has severe drawbacks: (1) traditional solution-based methods involving direct metallation of organic halides with metallic sodium suffer from a severe side reaction, Wurtz coupling, where the newly formed organosodium (RNa) reacts with unreacted organohalide (RX) to form the homocoupling product R‒R and sodium halide (NaX)30; (2) reactions often require a large excess of sodium lumps or unstable, expensive pre-activated sodium sources such as sodium dispersion13,17. To avoid Wurtz coupling, the mainstream method to prepare organosodiums is Li–Na exchange, by reacting organolithium (RLi) with sodium tert-butoxide (NaOtBu) to generate RNa and LiOtBu in a hydrocarbon solvent (usually hexanes), driven by the poor solubility of RNa in these solvents (see above)31. This Li–Na exchange still requires Li, which eliminates the sustainability merit of organosodiums. The need for rigorously dry and deoxygenated organic solvents and relatively complex setups involving inert-gas atmospheres presents additional practical drawbacks. Mechanochemical organic synthesis using ball-milling techniques has emerged as a new methodology to carry out organic synthesis under solvent-free conditions32,33,34,35,36,37,38,39,40,41,42,43. The advantages of this protocol include reduced solvent waste, fast reaction kinetics and operational simplicity under ambient conditions. Furthermore, recent studies have shown that vigorous mechanical agitation can activate zero-valent metals by removing the unreactive surface oxide layer and increasing the reactive surface area44,45,46,47,48,49,50,51,52,53. These mechanochemically activated, zero-valent metals readily undergo metal–surface reactions with organic halides to produce the corresponding organometallic species efficiently51. Surprisingly, by utilizing this strategy, organometallic compounds previously thought to be synthesizable only in solution have been successfully prepared under mechanochemical conditions, including Grignard reagents54,55,56,57, organozinc58,59,60, organocalcium61,62,63, organomanganese64,65, organobarium66 and organolithium reagents67. However, the direct mechanochemical synthesis of organosodium compounds has yet to be explored. Recently, mechanochemical activation of sodium lumps has been utilized for highly efficient ammonia-free Birch reduction and the preparation of a sodium anion complex (Fig. Building on these successful results, we envisioned that a mechanochemical protocol could offer a practical and efficient solution to address the challenges associated with conventional solution-based methods for the preparation and application of organosodium compounds in synthetic chemistry. In this Article, we report a mechanochemical generation of organosodium compounds from various organic halides and commercially available, cheap and stable sodium lumps and their application to organic synthesis (Fig. This protocol can be performed under ambient conditions without the need for pre-activated sodium metal, large amounts of anhydrous organic solvents or complex synthetic procedures that require precautions against moisture and strict temperature control. Furthermore, these rapid reactions complete within minutes. This is likely to be due to the mechanical activation of sodium metal in situ. The resulting organosodium species readily react with a wide range of electrophiles in a one-pot mechanochemical process. Nickel-catalysed direct cross-coupling of organosodium compounds with aryl halides also proceeded under mechanochemical conditions. Additionally, this mechanochemical strategy was applicable to the direct sodiation of poorly soluble aromatic halides as well as organic fluorides, which are unreactive towards sodiation under conventional solution-based conditions. Single-crystal X-ray diffraction (SCXRD) analysis and NMR spectroscopy studies were successfully performed to confirm the identity of the key mechanochemically generated organosodium intermediates. Overall, the present study provides a synthetic platform centred on organosodium compounds, which has the potential to replace conventional organolithium-based synthesis with a more sustainable, cost-effective and environmentally friendly approach. We initially attempted the generation of organosodium species through the reaction of bromobenzene (1a) with sodium lumps using a Retsch MM400 mixer mill (5 ml stainless-steel milling jar with one stainless-steel ball; ball diameter: 10 mm). The mineral oil on the sodium lumps was removed by wiping them with a paper towel, before the lumps were cut into pieces approximately 4–5 mm in width and depth. Subsequently, the sodium-metal pieces were weighed and introduced into the jar, followed by addition of 1a under atmospheric conditions. After optimization (see Supplementary Table 1 for details), we established the following mechanochemical protocol for generating organosodium compounds (Fig. 2): ball milling of 1a (2.0 equiv.) and sodium lumps (4.4 equiv.) in the presence of n-hexane (4.4 equiv.) as a liquid additive for efficient grinding was conducted for 5 min. The jar was opened in air and N-benzylideneaniline 2a (1.0 mmol) was added to the reaction mixture. The jar was then quickly closed (without any inert gas purging) and a second phase of ball milling was conducted for 5 min, yielding the desired amine 3a in excellent yield (87%). The scope of the aromatic halides and electrophiles using this one-pot, two-step mechanochemical protocol is shown in Fig. Phenylsodium could also be generated from chlorobenzene, and subsequent reaction with 2a produced 3a in 85% yield. Organosodium species were prepared from various alkylated aryl halides (1b–1f) and reacted with 2a smoothly to produce the corresponding amines (3b–3f) in moderate to excellent yields (50–91%). Biphenyl bromides (1g and 1h) were also readily metallated, and subsequent reaction with 2a afforded 3g and 3h in 53 and 61% yields, respectively. Aryl halides bearing trimethylsilyl, dimethylamino and methoxy groups were compatible with this reaction, and the corresponding amines (3i–3k) were obtained in moderate to high yields (34–73%). We tested the generation of organosodium compounds from sterically hindered aryl bromides such as mesityl bromide (1l) and 2-bromo-1,3,5-triisopropylbenzene (1m). Mesityl sodium was efficiently generated, and subsequent nucleophilic addition to 2a proceeded to give 3l in 90% yield. A more sterically hindered aryl halide 1m also underwent direct sodiation under mechanochemical conditions to afford the desired product 3m in 45% yield. N-(4-Methoxybenzylidene)aniline (2b) also reacted with the mechanochemically generated aryl sodium compound, 4-tolyl sodium, to give 3n in 66% yield. Conditions: 1 (2.0 mmol), Na (4.4 mmol), n-hexane (4.4 mmol) and electrophile (1.0 mmol) in a stainless-steel ball-milling jar (5 ml) with a stainless-steel ball (10 mm). Ball milling (30 Hz) was carried out. Isolated yields are reported as percentages. a15 min for first step. b15 min for first step and 30 min for second step. dFirst step: 1 (1.0 mmol), Na (2.2 mmol) and n-hexane (2.2 mmol) for 5 min. Second step: dry ice (excess) for 30 min. eFirst step: 1 (2.5 mmol), Na (5.0 mmol) and n-hexane (5.0 mmol) for 5 min. Second step: Ar–Cl (1.0 mmol), NiCl2(dppe) (0.10 mmol). iPr, isopropyl; tBu, tert-butyl; TMS, trimethylsilyl; pin, pinacol. In addition to imines as trapping reagents, aromatic and aliphatic aldehydes also smoothly reacted with mechanochemically generated organosodium compounds to afford the corresponding alcohols (3o–3s) in moderate to excellent yields (56–81%). Methyl benzoate could also be used as an electrophile and the desired tertiary alcohol 3t was obtained in 97% yield. Nucleophilic additions to aromatic ketones also proceeded rapidly, providing the corresponding alcohols (3u–3x) in moderate to excellent yields (42–80%). Additionally, various aromatic ketones (3y–3aa) were successfully prepared by reactions between organosodium compounds and Weinreb amides in good to excellent yields (65–85%). Morpholine amides were also reactive with the organosodium compounds, and the desired ketone 3ab and aldehyde 3ac were obtained in 53 and 67% yields, respectively. Reactions of dimethylphenyl silane (HSiMe2Ph) with the organosodium compounds bearing biphenyl and fluorene moieties provided the corresponding silylated products 3ad and 3ae in 92 and 89% yields, respectively. Mechanochemically generated phenylsodium could be trapped by dry ice, and benzoic acid 3af was obtained in 83% yield. The feasibility of mechanochemical nickel-catalysed cross-coupling reaction with organosodium compounds was also investigated70. Following catalyst optimization (see Supplementary Table 2 for details), we found that NiCl2(dppe) (dppe, 1,2-bis(diphenylphosphino)ethane) was an ideal catalyst for this reaction, and the cross-coupling between phenylsodium and 2-chlorolonapthalene gave the corresponding coupled product 3ag in 43% yield. In our substrate scope studies, side products such as homocoupled products of organic halides were not detected. Unfortunately, unlike aryl halides, alkyl halides proved difficult to use in one-pot, two-step reactions, and the desired products were not obtained. This was probably due to rapid decomposition of the reactive alkyl sodium intermediates formed during the metallation step. To address this limitation, we explored one-pot, one-step transformations in which the electrophile was introduced at the start of the reaction, enabling trapping of the reactive organosodium intermediate as soon as it was formed (Fig. Pleasingly, we found that employing this Barbier-type approach, a reaction between 1-chlorohexane, sodium lumps and imine 2a proceeded smoothly to provide the corresponding amine 3ah in 95% yield. Shorter and branched primary alkyl chlorides could also be used for this reaction, and the desired products 3ai and 3aj were obtained in 87 and 90% yields, respectively. Reactions using secondary and tertiary alkyl chlorides produced the corresponding amines 3ak and 3al in 63 and 18% yields, respectively. Fortunately, morpholine amides were also compatible under these strongly reducing conditions, enabling nucleophilic acyl substitution with n-butyl sodium to deliver the corresponding ketones (3am–3ap) in 42–66% yield. Nucleophilic addition to cyclopropyl phenyl ketone also readily proceeded and alcohol 3aq was obtained in 70% yield. Aldehyde electrophiles were also compatible, enabling access to secondary alcohol products (3ar–3at) in moderate to good yields (55–60%). Additionally, mechanochemical borylation of in situ-generated primary alkyl sodium species provided the corresponding boronic esters 3au and 3av in yields of 75 and 62%, respectively (in the latter case, 5% of Wurtz-type homocoupled product was detected as a minor side product). Conditions: 1 (2.0 mmol), Na (4.4 mmol), n-hexane (4.4 mmol) and electrophile (1.0 mmol) in a stainless-steel ball-milling jar (5 ml) with a stainless-steel ball (10 mm). Ball milling (30 Hz) was carried out. Isolated yields are reported as percentages. aAddition of n-hexane (4.4 mmol). b1 (1.0 mmol), Na (2.2–4.0 equiv.) and iPrO–B(pin) (2.0 mmol). iPr, isopropyl; tBu, tert-butyl; pin, pinacol. We also applied this operationally simple one-pot, one-step protocol to reactions using aryl sodium compounds (Fig. Under our optimized conditions, both 4-chlorotoluene and 4-tert-butylbromobenzene reacted with morpholine amides to furnish the corresponding ketones 3ab and 3aw in 76 and 74% yields, respectively. Nucleophilic additions of aryl sodium compounds to aldimine 2a and 4-tert-butylbenzaldehyde were also investigated, and the desired products 3b and 3o were successfully obtained in 77 and 40% yields, respectively. One-pot, one-step borylations were applied to 4-bromotoluene and 4-bromoanisole to deliver boronic esters 3ax and 3ay in 64 and 56% yields, respectively. These results are comparable to the corresponding one-pot, two-step reactions (Fig. To underline the practical utility of our developed mechanochemical protocol, we investigated its use in preparative-scale reactions (Fig. 4-Tolyl sodium was successfully prepared on a 12-mmol scale under mechanochemical conditions, and its nucleophilic addition to aldimine 2a gave 3b without any decrease in efficiency (96%, 1.582 g) compared to small-scale synthesis. Likewise, a one-pot, one-step reaction between 1-chlorobutane and 4-benzoylmorpholine was also carried out on a gram scale to furnish 3am in 69% yield. These results underscore the practical utility of this protocol. Isolated yields are reported as percentages. See Supplementary Section 4 for full details. To further expand the range of synthetically accessible organosodium compounds, we explored C–H sodiation of heteroaromatic arenes using a mechanochemically generated sodium amide (Fig. We were delighted to find that sodium 2,2,6,6-tetramethylpiperidide (NaTMP) could be generated by the reaction between phenylsodium and 2,2,6,6-tetramethylpiperidine (TMPH), and employed to deprotonate benzothiophene. Subsequent addition to aldimine 2a proceeded smoothly to afford 3az in 97% yield. Benzofuran and 1-methylindole were also deprotonated by mechanochemically generated NaTMP and the resulting organosodium compounds reacted with aldimine 2a to produce 3ba and 3bb in good yields. Aldehyde and Weinreb amide electrophiles could also be employed as electrophiles, delivering the corresponding alcohol (3bc) and ketone (3bd) products in yields of 93 and 85%, respectively. Isolated yields are reported as percentages. Conditions: 1a (1.0 mmol), Na (2.2 mmol), n-hexane (2.2 mmol), TMPH (1.0 mmol), 1 (1.0 mmol) and electrophile (0.5 mmol) in a stainless-steel ball-milling jar (5 ml) with a stainless-steel ball (10 mm). Ball milling (30 Hz) was carried out. See Supplementary Section 3F for full details. To highlight the synthetic utility of our developed mechanochemical protocol, we applied it to synthesize the drug molecule orphenadrine, an anticholinergic agent used to treat painful muscle spasms (Fig. After minor optimization, the generation of phenylsodium, followed by nucleophilic addition to 2-methylbenzaldehyde under mechanochemical conditions, afforded the intermediate 3q in 86% yield. Subsequent alkylation and reduction with LiAlH4 yielded orphenadrine in 85% yield over two steps. Isolated yields are reported as percentages. TMPH, 2,2,6,6-tetramethylpiperidine; THF, tetrahydrofuran; r.t., room temperature. See Supplementary Section 6 for full details. Due to the high basicity of organosodium species, the scope of available solvents is limited to simple hydrocarbon solvents. Therefore, aryl halides containing a large π-conjugated system, which are often sparingly soluble in such solvents, are difficult to directly sodiate under conventional solution-based conditions. For example, 1-bromo-3,5-diphenylbenzene (1be) is poorly soluble in n-hexane, and an attempted solution-state sodiation followed by addition to cyclopropyl phenyl ketone did not afford the desired product 3be (Fig. By contrast, we were delighted to find that our mechanochemical protocol was successfully able to deliver alcohol 3be in excellent yield (89%), highlighting the potential of this approach to access organosodium compounds that cannot be accessed via conventional solution-based reactions. a, Generation of organosodium compounds from a poorly soluble halide and nucleophilic addition to ketone. b, Generation of organosodium compounds from an organofluoride and nucleophilic addition to 2a. c, Various Na-meditated transformations of C–F bonds. Isolated yields are reported as percentages. Integrated 1H NMR spectroscopy yields are shown in parentheses. See Supplementary Section 5 for details. In a similar vein, we also explored the generation of organosodium species from organic fluorides via inert C–F bond cleavage (Fig. Direct metallation of organofluorine compounds is challenging because it involves breaking a strong C–F bond, and reports on the direct sodiation of fluoroarenes with unactivated sodium lumps are lacking73. Indeed, an attempted solution-phase (n-hexane) sodiation of fluorobenzene followed by addition of aldimine (2a) resulted in no product formation. On the other hand, we discovered that phenylsodium was rapidly generated from fluorobenzene in 5 min under mechanochemical conditions, and subsequent nucleophilic addition to aldimine 2a furnished 3a in 96% yield. We found that other sodium-meditated transformations via C–F bond activation were feasible under mechanochemical conditions (Fig. For example, sodiation of 4-fluorobiphenyl also proceeded smoothly, and nucleophilic addition to phenyl ethyl ketone produced 3u in 59% yield. Likewise, fluoroarenes bearing trimethylsilyl (TMS) and methyl groups could also be successfully sodiated, and nucleophilic substitutions with Weinreb amides afforded the corresponding ketones 3bf and 3bg in 76 and 53% yields, respectively. A Barbier-type reaction between [1,1'-biphenyl]-4-yl(morpholino)methanone and 1-fluoropentane afforded 3bh in 35% yield. 4-Tolyl sodium could be generated from 4-fluorotoluene and reacted with HSiMe2Ph to give the silylation product 3bi in 70% yield. Finally, to offer unambiguous evidence for the mechanochemical generation of the Na‒C bond during our reactions, we attempted to isolate and characterize a mechanochemically generated organosodium compound. For these structural studies, we selected a model aromatic organosodium bearing a lipophilic tert-butyl substituent, which we hoped would ensure sufficient hydrocarbon solubility to enable crystallization. Given that organosodium compounds are well known to form poorly soluble aggregates74 that could hamper their study by SCXRD, we also introduced a neutral amine ligand, namely N,N,Nʹ,Nʹ,Nʹ-pentamethyldiethylenetriamine (PMDTA), to break up these aggregates and form a soluble compound, allowing for crystallization. Accordingly, under an argon atmosphere, 3-chloro-tert-butylbenzene and 2.2 equiv. of Na metal were ball milled with 2.2 equiv. of n-hexane according to our standard optimized conditions (Fig. The crude product was treated with 0.6 equiv. of PMDTA solution in n-hexane followed by filtration to remove NaCl. Concentration followed by crystallization from n-hexane solution at −35 °C afforded a dimeric PMDTA-coordinated organosodium compound [Na(μC-3-tBu-C6H4)(κ3-N,Nʹ,Nʹ-PMDTA)]2 (4) (28% yield based on 1-(tert-butyl)-3-chlorobenzene). The single-crystal X-ray diffraction structure of 4 is shown in Fig. 8b, featuring two bridging phenyl groups, with Na‒C bond lengths within the range 2.61–2.65 Å. Solution state characterization of 4 was also performed by 1H, 13C and 23Na NMR spectroscopy in d12-cyclohexane. The 23Na NMR spectrum of 4 features a broad signal at –14.90 ppm (line width at half height LW1/2 approximately 10 ppm), which matches reported Na+ cation signals69. The 1H NMR spectrum of 4 clearly exhibits a characteristic pattern of four protons consistent with a 1,3-disubstituted phenyl group, and the PMDTA remains coordinated (Fig. The 13C NMR spectrum (d12-cyclohexane) of 4 features a Na‒Csp2 quaternary carbon signal at 195.1 ppm similar to other PMDTA-ligated aryl sodium complexes described in the literature75. The NMR spectra indicate that the SCXRD structure of 4 persists in solution in cyclohexane—this is particularly important in organo-alkali metal chemistry, where fast equilibria are often prevalent, such as ligand coordination–dissociation and deprotonation–reprotonation. The isolation and characterization of 4 unambiguously supports the formation of a Na‒C bond during the mechanochemical sodiation reactions in our methodology. a, Mechanochemical synthesis of [3-(tert-butyl)phenyl]sodium followed by ligation with PMDTA and crystallization to afford [Na(μC-3-tBu-C6H4)(κ3-N,Nʹ,Nʹ-PMDTA]2 (4). b, Single-crystal X-ray structure of 4. For clarity only the disordered components with the highest occupancy are depicted and hydrogen atoms are omitted. Key bond lengths (Å): Na1‒C3 2.616(2), Na1‒C3A 2.643(2), Na1‒N1 2.663(9), Na1‒N2 2.565(5), Na1‒N3 2.547(9). c, 1H NMR spectrum of 4 (d12-cyclohexane, 298 K). The colours in the 1H NMR spectrum indicate which signals correspond to the protons of compound 4. PMDTA, N,N,Nʹ,Nʹ,Nʹ-pentamethyldiethylenetriamine. Organosodium compounds have garnered substantial attention as promising, sustainable alternatives to organolithium reagents. However, their practical and widespread application in organic synthesis has been hindered by the lack of efficient generation methods using easy-to-handle sodium sources and their limited solubility in organic solvents. This study demonstrates that a mechanochemical protocol can overcome these challenges, enabling rapid synthesis of a wide range of organosodium compounds, including aryl, primary, secondary alkyl and benzylic examples from inexpensive, abundant sodium lumps and organic halides within minutes, without the need for large volumes of solvents or complex inert-gas techniques. The resulting mechanochemically generated organosodium compounds react smoothly with various electrophiles in a one-pot manner (either stepwise or Barbier-type conditions), negating any requirement to isolate the organosodium intermediates. The operational simplicity of the procedure, in particular the ability to proceed without exclusion of air is especially noteworthy, substantially lowering the barrier for synthetic chemists to make and apply these reagents. Notably, this method facilitated the direct sodiation of poorly soluble halides, which are typically unreactive to sodiation under solution-based conditions. Furthermore, the study revealed that the sodiation of organofluorides, which are generally inert in solution, could be achieved via C–F bond cleavage under mechanochemical conditions. These findings highlight the effectiveness of this mechanochemical approach in expanding the accessible scope of organosodium chemistry. Finally, the isolation and characterization of mechanochemically generated organosodium compounds by single-crystal X-ray analysis and NMR spectroscopy was performed, providing unambiguous support for the key organosodium species proposed in this methodology. Given the increasing demand for organosodium compounds, this efficient and straightforward mechanical strategy is expected to advance the development of sodium-based synthetic chemistry, contributing to more sustainable syntheses of value-added molecules. Sodium lumps (4.4 equiv.) were cut into small pieces (approximately 4–5 mm in width and depth) and weighed in air after wiping off the mineral oil on them with paper. These were then added into a milling jar (5 ml) with a ball (10 mm diameter). An organic halide (2.0 equiv.) and n-hexane (4.4 equiv.) were added to the jar. After the jar was closed without purging with inert gas, it was placed in the ball mill (Retsch MM400, 30 Hz). After grinding for 5 min, the jar was opened in air and charged with an electrophile (1.0 mmol, 1.0 equiv.) The jar was then closed without purging with inert gas, and was placed in the ball mill (Retsch MM400, 30 Hz). After grinding for 5 min, the reaction mixture was quenched with a saturated aqueous solution of NH4Cl and extracted with EtOAc (30 ml × 3). The solution was dried over MgSO4, filtered and evaporated to dryness under reduced pressure. The crude material was purified by flash chromatography (SiO2, typically EtOAc/n-hexane, typically 0–10:90) to give the corresponding product. The data that support the findings of this study are available in the Article and its Supplementary Information. For full characterization data, including NMR spectra of the new compounds and experimental details, see the Supplementary Information. The X-ray crystallographic coordinates for the structure reported in this study have been deposited at the Cambridge Crystallographic Data Centre (CCDC) under deposition number CCDC 2423000 (4). Copies of the data can be obtained free of charge via www.ccdc.cam.ac.uk/data_request/cif. (eds) The Chemistry of Organolithium Compounds (Wiley, 2004). Clayden, J. Organolithiums: Selectivity for Synthesis (Elsevier, 2002). Hodgson, D. M. (ed) Organolithiums in Enantioselective Synthesis (Springer, 2003). Wu, G. & Huang, M. Organolithium reagents in pharmaceutical asymmetric processes. Google Scholar Screttas, C. G., Steele, B. R., Micha-Screttas, M. & Heropoulos, G. A. Aryllithiums with increasing steric crowding and lipophilicity prepared from chlorides in diethyl ether. The first directly prepared room-temperature-stable dilithioarenes. Google Scholar Crockett, M. P., Aguirre, L. S., Jimenez, L. B., Hsu, H.-H. & Thomas, A. Preparation of highly reactive lithium metal dendrites for the synthesis of organolithium reagents. Google Scholar Tarascon, J.-M. Is lithium the new gold? Google Scholar Kim, T., Song, W., Son, D.-Y., Ono, L. K. & Qi, Y. Lithium-ion batteries: outlook on present, future, and hybridized technologies. Google Scholar Weiss, M. et al. Fast charging of lithium-ion batteries: a review of materials aspects. Google Scholar Wulandari, T., Fawcett, D., Majumder, S. B. & Poinern, G. E. J. Lithium-based batteries, history, current status, challenges, and future perspectives. Google Scholar Greenwood, N. N. & Earnshaw, A. (eds) Chemistry of the Elements 2nd edn (Butterworth-Heinemann, 1997). Weidmann, N., Ketels, M. & Knochel, P. Sodiation of arenes and heteroarenes in continuous flow. Google Scholar Asako, S., Nakajima, H. & Takai, K. Organosodium compounds for catalytic cross-coupling. Google Scholar Woltornist, R. A. et al. Structure, reactivity, and synthetic applications of sodium diisopropylamide. Google Scholar Fukazawa, M., Takahashi, F., Nogi, K., Sasamori, T. & Yorimitsu, H. Reductive difunctionalization of aryl alkenes with sodium metal and reduction-resistant alkoxy-substituted electrophiles. Google Scholar Asako, S., Takahashi, I., Nakajima, H., Ilies, L. & Takai, K. Halogen-sodium exchange enables efficient access to organosodium compounds. Google Scholar Harenberg, J. H. et al. (2-Ethylhexyl)sodium: A hexane-soluble reagent for Br/Na-exchanges and directed metalations in continuous flow. Google Scholar De, P. B., Asako, S. & Ilies, L. Recent advances in the use of sodium dispersion for organic synthesis. Google Scholar Bole, L. J., Tortajada, A. & Hevia, E. Enhancing metalating efficiency of the sodium amide NaTMP in arene borylation applications. Google Scholar & Hevia, E. Perdeuteration of arenes via hydrogen isotope exchange catalyzed by the superbasic sodium amide donor species NaTMP‧PMDETA. Google Scholar Anderson, D. E., Tortajada, A. & Hevia, E. Highly reactive hydrocarbon soluble alkylsodium reagents for benzylic aroylation of toluenes using Weinreb amides. Google Scholar Davison, N. et al. Divergent reaction patterns between organolithium and organosodium complexes and ligand-catalyzed ketone/aldehyde methylenation. Google Scholar Dilauro, G. et al. Introducing water and deep eutectic solvents in organosodium chemistry: Chemoselective nucleophilic functionalizations in air. Google Scholar On some new ethyl-compounds containing the alkalimetals. Google Scholar Benkeser, R. A., Foster, D. J., Sauve, D. M. & Nobis, J. F. Metalations with organosodium compounds. Google Scholar Schlosser, M. Organosodium and organopotassium compounds part Ⅰ: properties and reactions. Google Scholar Schlosser, M. Organosodium and organopotassium compounds part Ⅱ: Preparation and synthetic applications. Google Scholar & Mulvey, R. E. Alkali-metal mediation: diversity of applications in main-group organometallic chemistry. Google Scholar Anderson, D. E., Tortajada, A. & Hevia, E. New frontiers in organosodium chemistry as sustainable alternatives to organolithium reagents. Google Scholar Wurtz, A. Sur une nouvelle classe de radicaux organiques. Google Scholar Tortajada, A., Anderson, D. & Hevia, E. Gram-scale synthesis, isolation and characterisation of sodium organometallics: nBuNa and NaTMP. Google Scholar James, S. L. et al. Mechanochemistry: opportunities for new and cleaner synthesis. Google Scholar Do, J.-L. & Friščić, T. Mechanochemistry: a force of synthesis. Google Scholar Hernández, J. G. & Bolm, C. Altering product selectivity by mechanochemistry. Article PubMed Google Scholar Kubota, K. & Ito, H. Mechanochemical cross-coupling reactions. Tan, D. & García, F. Main group mechanochemistry: from curiosity to established protocols. Bolm, C. & Hernández, J. G. Mechanochemistry of gaseous reactants. Porcheddu, A., Colacino, E., De Luca, L. & Delog, F. Metal-mediated and metal-catalyzed reactions under mechanochemical conditions. & Hernández, J. G. Sustainability assessment of mechanochemistry by using the twelve principles of green chemistry. Bolt, R. R. A., Leitch, J. A., Jones, A. C., Nicholson, W. I. & Browne, D. L. Continuous flow mechanochemistry: reactive extrusion as an enabling technology in organic synthesis. Cuccu, F. et al. Mechanochemistry: New tools to navigate the uncharted territory of ‘impossible' reactions. Martinez, V., Stolar, T., Karadeniz, B., Brekalo, I. & Užarević, K. Advancing mechanochemical synthesis by combining milling with different energy sources. Alić, J., Schlegel, M.-C., Emmerling, F. & Stolar, T. Meeting the UN sustainable development goals with mechanochemistry. Jones, A. C., Leitch, J. A., Raby-Buck, S. E. & Browne, D. L. Mechanochemical techniques for the activation and use of zero-valent metals in synthesis. Davison, N. et al. A room-temperature-stable electride and its reactivity: Reductive benzene/pyridine couplings and solvent-free Birch reductions. Gao, Y., Kubota, K. & Ito, H. Mechanochemical approach for air-tolerant and extremely fast lithium-based Birch reductions in minutes. Fujishiro, K. et al. Lithium-mediated mechanochemical cyclodehydrogenation. Davison, N., Waddell, P. G. & Lu, E. Reduction of K+ or Li+ in the heterobimetallic electride K+[LiN(SiMe3)2]e−. Nallaparaju, J. V. et al. Mechanochemical Birch reduction with low alkaline earth metals. Kubota, K., Fukuzawa, Y., Kondo, K., Gao, Y. & Ito, H. Highly efficient and air-tolerant calcium-based Birch reduction using mechanochemistry. Bhattacharjee, D., Jana, S. K. & Maji, B. Solid-state mechanochemical Clemmensen reduction. Kubota, K., Nagao, A. & Ito, H. Solvent-free zinc meditated Béchamp reduction using mechanochemistry. Wada, S., Hayashi, N. & Suzuki, H. Noticeable facilitation of the bismuth-mediated Barbier-type allylation of aromatic carbonyl compounds under solvent-free conditions. Harrowfield, J. M., Hart, R. J. & Whitaker, C. R. Magnesium and aromatics: mechanically-induced Grignard and McMurry reactions. Takahashi, R. et al. Mechanochemical synthesis of magnesium-based carbon nucleophiles in air and their use in organic synthesis. Pfennig, V. S., Villella, R. C., Nikodemus, J. & Bolm, C. Mechanochemical Grignard reactions with gaseous CO2 and sodium methyl carbonate. Nallaparaju, J. V. et al. Mechanochemistry-amended Barbier reaction as an expedient alternative to Grignard synthesis. Cao, Q., Howard, J. L., Wheatley, E. & Browne, D. L. Mechanochemical activation of zinc and application to Negishi cross-coupling. Cao, Q., Stark, R. T., Fallis, I. & Browne, D. L. A ball-milling-enabled Reformatsky reaction. Yin, J., Stark, R. T., Faills, I. & Browne, D. L. A mechanochemical zinc-meditated Barbier-type allylation reaction under ball-milling conditions. Gao, P., Jiang, J., Maeda, S., Kubota, K. & Ito, H. Mechanochemically generated calcium-based heavy Grignard reagents and their application to carbon–carbon bond-forming reactions. Gao, P. et al. Direct arylation of alkyl fluorides using in situ mechanochemically generated calcium-based heavy Grignard reagents. Wang, X. et al. Direct arylation of gem-difluorostyrenes using in situ mechanochemically generated calcium-based heavy Grignard reagents. Nicholson, W. I. et al. Ball-milling-enabled reactivity of manganese metal. Takahashi, R., Gao, P., Kubota, K. & Ito, H. Mechanochemical protocol facilitates the generation of arylmanganese nucleophiles from unactivated manganese metal. Kubota, K., Kawamura, S., Jiang, J., Maeda, S. & Ito, H. Mechanochemical generation of aryl barium nucleophiles from unactivated barium metal. Kondo, K., Kubota, K. & Ito, H. Mechanochemical activation of metallic lithium for the generation and application of organolithium compounds in air. Kondo, K., Kubota, K. & Ito, H. Mechanochemistry enabling highly efficient Birch reduction using sodium lumps and D-(+)-glucose. Davison, N. et al. Mechanochemical synthesis of a sodium anion complex [Na+(2,2,2-cryptand)Na−] and studies of its reactivity: Two-electron and one-electron reductions. Inoue, T. et al. Cross-coupling polymerization of organosodium for polythiophene synthesis. Asako, S., Kodera, M., Nakajima, H. & Takai, K. Lithium-free synthesis of sodium 2,2,6,6,-tetramethylpiperidine and its synthetic applications. Ito, S., Takahashi, F. & Yorimitsu, H. Defluorinative diborasodiation of benzotrifluorides with bis(pinacolato)diboron and sodium. Lei, P. et al. A practical and chemoselective ammonia-free Birch reduction. Davison, N. & Lu, E. The quest for organo-alkali metal monomers: unscrambling the structure–reactivity relationship. A. et al. Donor-activated lithiation and sodiation of trifluoromethylbenzene: structural, spectroscopic, and theoretical insights. This work was supported by the Japan Society for the Promotion of Science (JSPS) via KAKENHI grants 22H00318 (H.I. ), by the JST via CREST grant JPMJCR19R1 (H.I.) and FOREST grant JPMJFR201I (K.K.) and by the Institute for Chemical Reaction Design and Discovery (ICReDD), which was established by the World Premier International Research Initiative (WPI), MEXT, Japan. thank the Leverhulme Trust for generous financial support via research grants RPG-2022-231 (E.L. and N.D.) and RPG-2023-159 (E.L., R.J.A. These authors contributed equally: Keisuke Kondo, Matthew Lowe. Division of Applied Chemistry, Graduate School of Engineering, Hokkaido University, Sapporo, Japan Keisuke Kondo, Koji Kubota & Hajime Ito Chemistry ‒ School of Natural and Environmental Sciences, Newcastle University, Newcastle upon Tyne, UK Matthew Lowe, Paul G. Waddell & Roly J. Armstrong School of Chemistry, University of Birmingham, Birmingham, UK Nathan Davison & Erli Lu Institute for Chemical Reaction Design and Discovery (WPI-ICReDD), Hokkaido University, Sapporo, Japan Koji Kubota & Hajime Ito Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar conceived and designed the study. All authors co-wrote the paper. performed chemical experiments and analysed the data. N.D. performed preliminary piloting experiments at an early stage of this project. collected and refined the single crystal structure of 4. All authors discussed the results and the manuscript. Correspondence to Roly J. Armstrong, Erli Lu, Koji Kubota or Hajime Ito. The authors declare no competing interests. Nature Synthesis thanks Andrea Porcheddu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Peter Seavill, in collaboration with the Nature Synthesis team. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Sections 1–15, Figs. 1–10, Tables 1–3 and Experimental details. X-ray crystallographic data for compound 4, CCDC 2423000. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/. Reprints and permissions Kondo, K., Lowe, M., Davison, N. et al. Mechanochemical synthesis of organosodium compounds through direct sodiation of organic halides. Version of record: 08 December 2025 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative © 2025 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Controlled multimode wave propagation can enable more space-efficient photonic processors than architectures based on discrete components connected by single-mode waveguides. Instead of defining discrete elements, one can sculpt the continuous substrate of a photonic processor to perform computations through multimode interference in two dimensions. Here we designed and demonstrated a device with a refractive index that can be rapidly reprogrammed across space, allowing arbitrary control of wave propagation. The device, a two-dimensional programmable waveguide, uses parallel electro-optic modulation of the refractive index of a slab waveguide with about 104 programmable spatial degrees of freedom. We implemented neural network inference on benchmark tasks with up to 49-dimensional vectors in a single pass, without digital pre-processing or post-processing. Theoretical and numerical analyses further indicated that two-dimensional programmable waveguides may offer not only a constant-factor reduction in device area but also a scaling benefit, with the area required growing as N1.5 rather than N2. Deep neural networks (DNNs) have gained widespread adoption across many domains ranging from computer vision to natural language processing1. The size of DNN models has been increasing exponentially over the past decade, leading to exponentially increasing energy costs for running them. Limits to energy costs now impose a practical constraint on how large models can be2, strongly motivating the exploration of alternative and energy-efficient computing approaches for executing DNNs, with a computational cost that is typically dominated by that of matrix–vector multiplications (MVMs). Optical neural networks, particularly those integrated on photonic chips that specialize in performing MVMs with optics instead of electronics, are one promising candidate approach3,4,5,6,7,8,9. The dominant paradigm for designing integrated photonic neural networks is to construct networks of discrete, programmable photonic components, such as Mach–Zehnder interferometers (MZIs), microring resonators or phase-change memory cells, connected by single-mode waveguides10. However, the maximum vector size, N, supported by chips using this approach has so far been restricted (Supplementary Table 1) to sizes far below what is necessary for optics to deliver an energy-efficiency advantage (N ≳ 1,000)11,12,13,14. The scale of such chips has been limited by at least two factors. Second, individual optical components are rather bulky owing to the comparably large optical wavelength and the often weak programmability of optical materials. On top of that, substantial portions of the chip's area need to be dedicated to sprawling interconnection regions comprising well-isolated waveguides. We could achieve far greater spatial efficiency if we treated the entire chip as a blank slate with refractive index distribution, n(x, z), as a function of space, which we could arbitrarily and reprogrammably sculpt, instead of building the integrated photonic neural network from discrete components15,16,17,18,19,20. This study addressed a central challenge: how to make a photonic chip with a programmable refractive index distribution, ideally without the integration complexity of introducing electronic wiring. In conventional nanophotonic chips, n(x, z) is controlled by etching away material in lithographically defined regions and is fixed at fabrication time. Although inverse-designed chips21 realizing MVMs with fixed matrices can be made19, we generally would like to be able to program the matrix. Photorefractive crystals were explored several decades ago as a means to implement programmable linear operations with slab waveguides22,23 but fell out of favour because the small achievable refractive index modulation (10−4) meant that even centimetre-scale waveguides were unable to perform large-scale operations. Additionally, phase-change materials have recently been demonstrated to realize arbitrary refractive index distributions24,25 but suffer from a limited number of rewrite cycles (4,000 in ref. 26) and high loss (greater than 2.8 dB mm−1 in refs. 25,26). Here we introduce a photonic chip with a waveguide that is fully programmable in two dimensions: a two-dimensional (2D) programmable waveguide. The chip uses massively parallel electro-optic modulation to program n(x, z) across ~10,000 individual regions of a lithium niobate slab waveguide. We trained multimode photonic structures within the chip that perform neural network inference (Fig. 1). The structures realized by our 2D-programmable waveguide are similar to inverse-designed nanophotonic devices19,21; they are computer-optimized 2D metastructures that control multimode wave propagation. A distinguishing feature of our device is its programmability, setting it apart from typical inverse-designed photonic devices, which are fixed after manufacturing. We achieve programmability optically, decoupling the electronic wiring for programming from the photonic chip; a pattern of light shone on top of our device creates a spatially varying refractive index modulation Δn(x, z) in the slab waveguide. This is achieved by using the principle of photoconductive gain27,28 to induce a refractive index modulation via the strong electro-optic effect in lithium niobate. In our study, we show how we can train the refractive index distribution so that the complex wave propagation through the device performs a desired neural network inference (Fig. a, The fundamental unit of an artificial neural network (a layer) transforms an input vector into an output vector via a trainable matrix multiplication. b, Analogous to a neural network layer, the 2D-programmable waveguide linearly transforms an input optical field into an output optical field via wave propagation through a lithium niobate slab waveguide with 2D refractive index modulation Δn(x, z) that can be continuously and arbitrarily programmed (up to practical limits on resolution and maximum modulation; Supplementary Information sections 1C,D). This refractive index modulation, which is directly set by the illumination pattern projected onto the device (shown in green), is trained so that wave propagation through the waveguide performs machine learning tasks (handwritten-digit classification shown here as an example). c, Simulated wave intensity in the slab waveguide, which shows that the neural network computation is performed with complex multimode wave propagation. Figure 1 shows a schematic of the machine learning process using the device. In machine learning, data are often encoded in vector form, \({\bf{x}}_{l}\), and processed via programmable matrices W and activation functions f in ‘layers': \({\bf{x}}_{l+1}=f(W{\bf{x}}_{l})\) (illustrated in Fig. 1a). We amplitude-encoded the machine learning input data \({\bf{x}}_{l}\) into the one-dimensional (1D) input optical field distribution, E(x, z = 0) (Fig. 1b), which serves as the initial condition for programmable wave propagation described by the following partial differential equation: A solution of the equation is shown in Fig. The refractive index distribution of the slab waveguide, n(x, z) = n0 + Δn(x, z), has two contributions: a spatially uniform part, n0, which is the refractive index of the waveguide when no programming light is impinging on it, and a programmable part that is induced by electro-optic modulation, Δn(x, z), loosely corresponding to the programmable matrix weights W. After the optical field has propagated through the device, we measured its intensity, Iout(x) ∝ ∣E(x, z = L)∣2, at the output facet and binned it to produce the output vector, \({\bf{x}}_{l+1}\), for the machine learning task. Figure 2 shows how patterns of light program the refractive index modulation, Δn(x, z). First, the device converts the projected pattern of light into a spatially varying quasi-static electric field. Second, the electro-optic waveguide changes its refractive index Δn(x, z) in response to this electric field. Our device is an electro-optic waveguide coated by a photoconductive film and sandwiched between a pair of electrodes with an oscillating bias voltage applied across them. The lithium niobate waveguide locally changes its refractive index via the Pockels effect, owing to the spatially varying bias electric field. In our device, the largest refractive index modulation is approximately 10−3, limited by the geometry of the material stack and a safety margin to prevent dielectric breakdown (Supplementary Information section 1C). The refractive index modulation in our device can take on continuous values by continuously varying the intensity of the projected pattern. a, The 2D-programmable waveguide consists of a nanophotonic stack of four layers: a conductive silicon substrate that doubles as the ground electrode, a Z-cut lithium niobate (in red) slab waveguide with silicon dioxide cladding (in white), a photoconductive layer for optical control of the refractive index and a gold electrode. b,c, Electrical circuit models of the 2D-programmable waveguide in regions with (c) and without (b) illumination. There is a voltage division between the photoconductor and the slab waveguide, with impedances, Zpc and Zwg, respectively. e, Experimental realization of a Y-branch splitter on the 2D-programmable waveguide, which splits the input light into two equal output beams. The projected pattern (in green) directly corresponds to the induced refractive index modulation (in grey). We projected illumination patterns across a 9 mm × 1 mm area, achieving a pixel resolution of 9 μm × 9 μm. (Limits to the spatial resolution are discussed in Supplementary Information section 1C.) This configuration enabled us to control the refractive index distribution, n(x, z), with 10,000 degrees of freedom (see Supplementary Information section 3A for our calculation of the parameter count) and update the entire distribution at a rate of 3 Hz. To maximize the refractive index modulation, we set the amplitude of the oscillating bias voltage across the electrodes to be up to 1,000 V. Given that complementary metal–oxide–semiconductor electrode backplanes can only support spatially programmable voltages of around 10 V, our use of photoconductive gain was essential for achieving large electro-optic modulation. This approach allowed us to apply a large voltage to a single unpatterned electrode and realize controllable high voltages at virtual electrodes via the patterned illumination27,28. Owing to the large impedance of the device, less than 1 mW of electrical power was dissipated across its active area at the highest voltages. To illustrate the operating principle of our device, we projected a pattern in the shape of a Y-branch splitter onto the 2D-programmable waveguide (Fig. 2a). The refractive index modulation was approximately proportional to the projected pattern, Δn(x, z) ∝ I(x, z) (Fig. 2e), up to spatial smoothing and a weak nonlinearity owing to voltage division. We coupled a single input Gaussian beam into the device using a beam shaper and measured the intensity of the output light with a camera, as shown in Fig. 2e. Because we used an oscillating bias voltage to drive the device, the induced refractive index pattern also oscillated in time. Therefore, measurements of the output intensity (and of the optical inputs, if not continuous wave) must coincide with the driving voltage peaking. We discuss approaches to overcome this limitation and further details on the experimental set-up in Methods. We next applied the 2D-programmable waveguide to perform machine learning tasks: vowel classification29 and Modified National Institute of Standards and Technology (MNIST) handwritten-digit classification30. Both tasks are used as benchmarks in studies of similar on-chip optical neural networks, providing useful points of comparison3,6,20. The vowel classification dataset29 comprises formant frequencies extracted from audio recordings of spoken vowels by various speakers. The task was to predict which of the seven vowels is spoken, given a 12-dimensional input vector of formant frequencies. We divided the dataset into a training set and a test set, comprising 196 (75% of the dataset) and 63 (25%) samples, respectively. Figure 3 presents our experimental results on performing vowel classification with the 2D-programmable waveguide. For readout, we measured the intensity at the output facet with a camera and binned the camera pixels into seven different regions, with each region corresponding to a specific vowel. Thus, as shown in the simulated intensity distribution, ∣E(x, z)∣2, in Fig. 3b, the device learned to use complex multimode wave propagation to direct most power towards the region corresponding to the correct vowel. Further details on output decoding and the overall computational model of our optical neural network demonstrations are provided in Supplementary Information section 5A. a, Overview of the approach. The task involved predicting a spoken vowel, here ‘er', from a 12-dimensional input vector representing formant frequencies extracted from audio recordings. The 2D-programmable waveguide was trained to take in this input vector and output a seven-dimensional vector with a one-hot encoding format that indicates the predicted vowel. b, Left: the input vector was amplitude-encoded into 12 Gaussian spatial modes to produce the initial optical field distribution. Right: the experimentally measured output intensity was binned by calculating the total power within equally sized spatial bins to produce the seven-dimensional output vector. c, Illustration of physics-aware training, a hybrid in situ–in silico backpropagation algorithm, which we used to train the parameters of the 2D-programmable waveguide. d, Test accuracy as a function of epoch. e, Evolution of the trainable parameters (projected patterns) at different stages of training. The refractive index distribution to implement vowel classification was learned using physics-aware training31, a modified backpropagation algorithm (Fig. 3c). The hybrid in situ–in silico nature of the algorithm allows for efficient training even in the presence of both imperfect models and experimental noise (Supplementary Information section 3C). Initially, a purely physics-based model (using equation (1)) provided qualitative but not quantitative agreement with the experimental results. The remaining discrepancies were largely removed with data-driven refinements to the physics-based model (Supplementary Information section 4). Using physics-aware training, we trained the 2D-programmable waveguide for a total of 300 epochs, which took approximately 1 h on the experimental set-up (Fig. 3d). Figure 3e shows the evolution of the initially uniform illumination pattern into a complex pattern that resembles the refractive index distributions found in inverse-designed photonic devices. Figure 4 presents the experimental results on MNIST handwritten-digit classification. We divided the MNIST dataset in the standard manner into 60,000 training images and 10,000 test images. We downsampled each MNIST image to 7-by-7 pixels and then flattened it to a 49-dimensional input vector. a, We performed MNIST handwritten-digit classification with the 2D-programmable waveguide. Each image from the MNIST dataset was electronically downsampled and reshaped to a 49-dimensional vector. We trained the device to perform machine learning on this high-dimensional input vector with the same procedure as the vowel classification task (Fig. 3). b, The confusion matrix was derived from evaluations on the test dataset comprising 10,000 images. After ten epochs of training, the system achieved 86% accuracy on the test dataset. To train the refractive index distribution to perform MNIST classification, we followed the same procedure used for the vowel classification task (Fig. 3). The 2D-programmable waveguide processed the 49-dimensional input vector to produce a ten-dimensional output vector that corresponds to the ten possible digits (more details in Supplementary Information section 5C). 4b, the system achieved 86% accuracy on the test dataset after ten epochs of training, which took about 10 h on the experimental set-up. This falls 4% short of the 90% accuracy that a one-layer digital neural network achieves on this downsampled MNIST classification task, likely because of imperfect modelling and experimental drifts. Nevertheless, this result demonstrates that complex wave propagation in our device can be harnessed to perform computations comparable to that of a single-layer neural network with a 49 × 10 matrix of trainable parameters. We turn to a discussion of the theoretical size scaling of different on-chip photonic processors. Integrated photonics can be used to implement matrices of size N × N using circuits of width much wider than ~Nλ0 (N well-isolated waveguides, where λ0 is the free-space wavelength) and length greater than ~Nλ0/Δnprog (N π-phase shifters)32,33. Because the programmable refractive index Δnprog is often very small, this limits circuits to be either very long or operate on low-dimensional inputs. One might intuitively expect that the length Lz of 2D-programmable waveguides would also need to scale as ~Nλ0/Δnprog, just as it does for the aforementioned circuit approaches for universal linear transformations on chip32. In this section, we present the analytical and numerical results showing that, beyond the constant-factor improvement in spatial footprint by avoiding single-mode waveguides, 2D-programmable waveguides may offer a different size scaling. Our analytical argument, presented in detail in Supplementary Information section 8A,B, investigates the amount of phase shift necessary to optically implement a given unitary transformation of dimension N in a programmable multimode waveguide. (More precisely, for any given unitary, one can always find a generator with elements that have a root-mean-square (r.m.s.) This suggests an exciting possibility. High-dimensional optical matrix–vector multipliers based on interference may require much less propagation distance through phase shifters than commonly assumed. This property of unitary matrices has recently34 been exploited to show that in three-MZI meshes, the average length of phase shifters can scale as \(1/\sqrt{N}\), but universal MZI meshes by construction have a circuit depth that scales as N, independent of how short the individual phase-shifter elements are. However, in 2D-programmable waveguides, this insight can potentially lead to large practical benefits because the devices effectively consist entirely of phase shifters, with nothing else constraining the total system length. The intuitive core of the mathematical argument is that if one needed a full π-phase shift per phase shifter, the required device length would scale as Nλ0/Δnprog, but because the required phase shift goes as \(\pi /\sqrt{N}\), the total required length scales as \(N{\lambda }_{0}/(\sqrt{N}\Delta {n}_{{\rm{prog}}})=\sqrt{N}{\lambda }_{0}/\Delta {n}_{{\rm{prog}}}\). We analytically found a refractive index distribution, Δnprog(x, z), with a magnitude (or, equivalently, length) that scales only as \(\sqrt{N}\) (as measured by the r.m.s. Our construction is similar to the one presented by Larocque and Englund17 in that it couples the modes of a multimode waveguide but crucially differs by implementing parallel global couplings rather than a sequence of pairwise mode couplings, thereby achieving better scaling. Our analytical argument relies on strong approximations. We used coupled-mode theory under a rotating wave approximation to calculate the propagation of unidirectional scalar waves in a perturbed multimode waveguide. To validate our theory, we present the results from numerical simulations that make far fewer assumptions. We simulated unidirectional scalar waves in a perturbed multimode waveguide (Supplementary Information section 8D). Our analytical argument only shows that the r.m.s. of the refractive index distribution scales as \(\sim \sqrt{N}\). However, in practical devices, it is usually the maximum value of the refractive index change that is limited. Our simulations show that even with a strictly imposed maximum value of the programmable refractive index strength, unitaries can be implemented accurately over a propagation distance that scales as \(\sim \sqrt{N}\). We emphasize that our analytical argument permits completely general dense unitary matrices. This length scaling is surprising in part because fundamental geometric considerations in linear optical devices suggest that the length required to perform arbitrary operations on N modes is proportional to Nλ0 (refs. 35,36). Our theory, showing that \({L}_{z} \sim \sqrt{N}{\lambda }_{0}/\Delta {n}_{{\rm{prog}}}\) is sufficient to perform universal optical operations on N modes, implies a different length scaling but does not contradict the bound by Miller35 unless N is very large, often ≫1,000 using common values (Supplementary Information section 8D). Therefore, our result suggests that there is a large practical regime in which the better-than-linear length scaling can be exploited to create universal optical processors with modest programmability or short propagation distances. However, we note that the length-scaling argument does not directly apply to the device used in our experiments because our scaling result assumes guided modes in the x direction, whereas our experiments were performed in a slab waveguide wide enough that light was effectively unguided. It is nonetheless the case that one could, as far as we can tell without having explicitly performed these experiments, realize devices and experiments that satisfy the assumptions of the theory (see discussion in Supplementary Information section 8D). For example, we simulated a multimode waveguide with a modest width of 0.3 mm, length of 5 mm and a refractive index programmability no larger than Δnprog = 5 × 10−3 with 500-nm resolution, and showed that such a waveguide can realize a 200 × 200-dimensional unitary with high fidelity (Fig. A device measuring 1.5 mm in width and 11 mm in length, with the same programmable refractive index magnitude and resolution, should be able to realize arbitrary unitaries with dimensions as high as 1,000 × 1,000. a, We considered a weak refractive index profile Δnprog(x, z) embedded into a step-index multimode waveguide with profile nwg(x). As light propagated through the perturbed waveguide, its N eigenmodes interacted with each other, as described by the coupled-mode theory. b, Vectors \(\bf{a}\) are encoded in the modal amplitudes of the input electric field. Given a unitary matrix Utarget, we analytically determined the refractive index perturbation Δnprog(x, z) embedded in the step-index waveguide (Supplementary Information section 8B) or via in silico inverse design (Supplementary Information section 8C). We verified that after propagation through the multimode waveguide, the modal amplitudes transformed according to Urealized ≈ Utarget. The separation between unitaries realized with high versus low fidelity is well described by the line \({L}_{z}=\sqrt{N}{\lambda }_{0}/\Delta {n}_{\max }\). We introduced and demonstrated a 2D-programmable photonic processor comprising a lithium niobate slab waveguide with a refractive index distribution, n(x, z), that can be continuously programmed. The device design enables programming by parallel electro-optic modulation with approximately 10,000 degrees of freedom. We used our chip to perform neural network inference by training the refractive index distribution and consequently the multimode wave propagation through the chip. To train the device, we developed a physics-based model of the behaviour of the chip, along with a data-driven refinement, allowing the model to be sufficiently accurate that it supports backpropagation-based training31. The predominant approach to building integrated photonic neural networks is to fabricate large arrays of discrete components connected by single-mode waveguides10. In contrast, we adopted the conceptual approach of using wave propagation in distributed spatial modes15,16,17,18,19,23,37 and experimentally validated the theoretical predictions16,17,37 that this approach will be more space-efficient. Our prototype chip was able to perform neural network inference with input vectors of dimension up to 49, which is larger than the capability of the neural network photonic chips reported in previous studies3,6,8,9,20,38,39,40 and more space efficient than any of these chips based on networks of discrete components (see Supplementary Information section 6 for a detailed comparison). This large input dimension enabled us to use our chip to perform MNIST handwritten-digit classification with a single pass through the chip and without using any digital electronic parameters. One of the most promising applications of our approach is in reducing the energy cost of neural network inference, which remains (and is likely to remain) dominated by linear MVM. There exists a breakeven point beyond which optical devices could outperform electronic hardware on this metric, owing to the more favourable energy scaling with dimension N (optics: E ~ N; electronics: E ~ N2) (refs. 11,12,13,14). Reaching this regime requires high-dimensional operations. At low N, the overhead cost from analogue-to-digital conversion and optical–electronic transduction outweighs the benefits, with estimates placing the breakeven point around N = 1,000, far beyond what is currently possible on a single chip. We derived a theoretical scaling law (Supplementary Information section 8) describing how the dimension of possible MVMs in a 2D-programmable waveguide scales with the device dimensions and the refractive index change. Surprisingly, the device length only needs to scale as \(\sqrt{N}\), better than the most common approaches of designing photonic circuits32,33. This result may enable all-optical matrix–vector multipliers with a dimension exceeding the breakeven point of energy efficiency. The development of such devices would make hybrid neural network architectures, in which analogue optics performs the linear operations and electronic circuits implement the nonlinearities, energy-competitive, changing the energy scaling of neural network inference. To conclude, we believe that our device concept, with its ability to programmably control multimode wave propagation, may create new opportunities in the broader fields of optical computing and optical information processing10,14,41. Although our study has focused on machine learning, our device could also be used to solve integral equations42 and combinatorial optimization problems43. More broadly, our chip is essentially an arbitrary (passive) photonic device that can be reconfigured on demand. Any photonic device that can be specified as an inhomogeneous refractive index distribution can be realized. Such devices can even be learned directly and effectively by performing inverse design21 but in situ in real time. Our concept will potentially enable the development of reprogrammable photonic simulators supporting new studies of bound states in the continuum44 and topological photonics45 and applications in engineering. It may ultimately even be possible to make a device that combines programmable linear wave propagation (this study), programmable nonlinear wave propagation15,18,46,47 (a natural extension of this study to having programmable χ(2)(x, z) (ref. 48)) and programmable gain/loss (demonstrated in ref. 20), giving rise to a reconfigurable on-chip platform capable of realizing almost every functionality we have in free-space optics. As shown in Extended Data Fig. 1, we started our fabrication processes from a thin-film lithium niobate wafer purchased from NANOLN. It was a p-type silicon wafer with a substrate conductivity of 0.01–0.05 Ω cm, 2 μm of silicon dioxide deposited via plasma-enhanced chemical vapour deposition (PECVD) and 700 nm of Z-cut MgO-doped lithium niobate that was wafer-bonded with the ion-cut technique. We diced small pieces from the wafer using a DISCO Dicing Saw for further processing. We deposited an additional 1 μm of silicon dioxide via PECVD as a cladding, followed by another deposition of 4 μm of silicon-rich silicon nitride, which was the photoconductive layer, via PECVD. We alternated pulses of high-frequency and low-frequency power during deposition to minimize film stress using 160-W low-frequency pulses for 12 s and 200-W high-frequency pulses for 8 s. Next, we evaporated electrodes onto the chip using a CVC SC4500 E-gun Evaporation System. We first evaporated 10 nm of titanium as an adhesion layer and then 5 nm of gold. The tape acted as a mask, preventing deposition closer than around 1 mm to the edges, thereby increasing the path length between the top electrode and substrate through air. To minimize coupling losses into the waveguide, we used an Allied MultiPrep Polisher to polish the waveguide facets. We polished using silicon carbide paper of successively finer grain size, starting at 3 μm and then moving to 1 μm and 0.5 μm. As shown in Supplementary Fig. 1, we used the transverse magnetic mode of the slab waveguide, with an optical electric field that is also oriented in the y direction. The thickness of the lithium niobate layer was chosen for single-mode operation (Supplementary Information section 1A). To maximize the refractive index modulation, it is beneficial to have a thicker photoconductor and a thinner silicon dioxide cladding (Supplementary Information section 1B). The silicon dioxide cladding was chosen to be sufficiently thick to ensure low propagation loss. Thus, we balanced these tradeoffs to arrive at the device geometry shown in Extended Data Fig. 1. The device has a propagation loss of <1 dB cm−1 at a wavelength of 1,550 nm (Supplementary Information section 1E). To maximize the refractive index contrast of the slab waveguide between the bright (illuminated) and dark regions, it is important to design the photoconductor to have high dark resistance and low bright resistance. We characterized the refractive index modulation as a function of the intensity of the projected pattern with an off-axis holography set-up (Supplementary Information section 1C). The maximum refractive index modulation that we achieved in this study was approximately 10−3. We show in Supplementary Fig. 3 that this can be increased to beyond 4 × 10−3 by using a photoconductor layer that is twice as thick (8 μm) and by further optimizing the photoconductive properties. In Supplementary Information section 7B, we also discuss how the refractive index modulation can be further increased by switching to a different material for the waveguide core. As shown in Extended Data Fig. 2, the experimental set-up can be roughly divided into five units: (1) an optical beam shaper to create spatially varying 1D electric field inputs for the 2D-programmable waveguide; (2) a projector to create a programmable illumination pattern that controls the refractive index distribution inside the waveguide; (3) a butt-coupling set-up to couple light in and out of the 2D-programmable waveguide; (4) a high-voltage source to apply an oscillating bias voltage across the electrodes of the 2D-programmable waveguide; and (5) a camera to measure the intensity of the output beam. We note that the experimental set-up relies on more free-space optical components than usual for an integrated photonics experiment. This is a direct consequence of our decision to keep the fabrication of the device simple, without lithographically defined structures for this initial proof-of-concept demonstration. We envision that a more compact, fully integrated version of the 2D-programmable waveguide could be built by integrating on-chip lithium niobate modulators and detectors and a micro-light-emitting-diode display (Supplementary Information section 7D). In this section, we provide an overview of the key components and functionalities of the experimental set-up. For a more detailed description, including photographs of the optical set-up and specifics on the components, such as part numbers and manufacturers, see Supplementary Information section 2. The free-space beam shaper allows for the realization of arbitrary input optical fields E(x, z = 0) up to a spatial resolution of 2 μm and over a distance of 600 μm. In this experiment, we used the beam shaper to create both simple input fields, such as a single Gaussian beam for the Y-branch splitter demonstration, and more complex input fields for the machine learning demonstrations. This flexibility also enabled us to freely vary the encoding of input vectors into the optical field. For instance, we varied the width of the input modes and adjusted their spacing, which was tailored to each machine learning task. Finally, because the beam shaper is capable of shaping both the amplitude and phase of the input field, it was also used to calibrate the 2D-programmable waveguide (Supplementary Information section 4B). The design of the beam shaper that we built closely follows that of a previous study49, which also programmably shapes the input light coupled into slab waveguides. The core working principle of the beam shaper is to create spatially varying phase gratings on a 2D-phase spatial light modulator50 (SLM; Meadowlark Optics UHSP1K-850-1650-PC8). We varied the amplitude and relative positions of these phase gratings to control the input optical field E(x, z = 0). A lens after the SLM performed a Fourier transform that separated the diffraction maxima of the phase gratings, and a spatial filter selected the first-order diffraction maximum. Finally, an optical relay system (comprising lenses 2 and 3, as shown in Extended Data Fig. The projector set-up was designed to create a high-resolution programmable illumination pattern over a large field of view. We used a digital micromirror device (DMD; Vialux V-7000) with a resolution of 1,024 × 768 pixels and a pixel pitch of 13.7 μm. The DMD was illuminated with green light (525 nm) from an LED. The focal length of the tube lenses was chosen to demagnify the image of the DMD by a factor of 1.5, resulting in a projected pattern on the surface of the 2D-programmable waveguide with dimensions of 9.1 mm × 6.8 mm, with each individual pixel of the projected pattern measuring 9 μm × 9 μm. Because the complex wave propagation spans a distance of 1 mm in the x direction, in practice, we used only a 9.1 mm × 1 mm region of the projected pattern to control the wave propagation in the 2D-programmable waveguide. Finally, although the DMD provides only binary modulation, we achieved continuous refractive index modulation by applying pulse-width modulation to the illumination pattern. This is feasible because the DMD can be switched on and off at a rate of 20 kHz, much faster than the resistance–capacitance time constant of the device, which is about 10 Hz. To maximize the electro-optic effect in lithium niobate, we used high voltages of about 1 kV. We created sinusoidal voltages with an arbitrary function generator and amplified the voltage with a Trek 2220 high-voltage amplifier, which has a voltage gain of 200× and is capable of outputting voltages of up to 2 kV. We electrically contacted the device using high-voltage-rated probe arms with BeCu probe tips; one probe tip was put in contact with the gold electrode on top of the device, whereas a grounded probe tip touched the silicon substrate (Supplementary Fig. We used an a.c. frequency of 10 Hz for the experiments shown in Fig. 2 and 26 Hz for the experiments shown in Figs. Future modifications can enable d.c. operation, such as using an alternative cladding material that is more conductive or by increasing the dark resistivity of the photoconductor (Supplementary Information section 1B). To measure the output of the computation performed by our device, we imaged the output facet of the device with an infrared camera (Allied Vision Goldeye CL-033). We built a 4f relay with a magnification factor of 5.3, allowing us to image the intensity distribution at the output facet, Icamera(x, y), with a resolution of 2.8 μm per pixel and a field of view of 1.7 mm in the x direction. We defined a small range of y to be the region of interest and integrated the intensity over this range to obtain the 1D intensity output of the 2D-programmable waveguide: \({I}_{{\rm{out}}}(x)=| E(x,z=L){| }^{2}=\mathop{\int}\nolimits_{{y}_{\min }}^{{y}_{\max }}{I}_{{\rm{camera}}}(x,y){\rm{d}}y\). We used an exposure time on the order of 500 μs, chosen to be much shorter than the period of the a.c. voltage, which was approximately 40 ms. All data generated during this work are available via Zenodo at https://doi.org/10.5281/zenodo.10775721 (ref. An expandable demonstration code for simulating wave propagation through programmable waveguides is available at https://github.com/mcmahon-lab/2D-programmable-waveguide. The code used for this work is available via Zenodo at https://doi.org/10.5281/zenodo.10775721 (ref. Carbon emissions and large neural network training. Deep learning with coherent nanophotonic circuits. Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H. & Pernice, W. H. All-optical spiking neurosynaptic networks with self-learning capabilities. Parallel convolutional processing using an integrated photonic tensor core. Huang, C. et al. A silicon photonic–electronic neural network for fibre nonlinearity compensation. & Aflatouni, F. An on-chip photonic deep neural network for image classification. Bandyopadhyay, S. et al. Single-chip photonic deep neural network with forward-only training. Shastri, B. J. et al. Photonics for artificial intelligence and neuromorphic computing. Hamerly, R., Bernstein, L., Sludds, A., Soljačić, M. & Englund, D. Large-scale optical neural networks based on photoelectric multiplication. Nahmias, M. A. et al. Photonic multiply-accumulate operations for neural networks. Anderson, M. G., Ma, S.-Y., Wang, T., Wright, L. G. & McMahon, P. L. Optical transformers. McMahon, P. L. The physics of optical computing. A. D., Minkov, M. & Fan, S. Wave physics as an analog recurrent neural network. Khoram, E. et al. Nanophotonic media for artificial neural inference. Larocque, H. & Englund, D. Universal linear optics by programmable multimode interference. Nakajima, M., Tanaka, K. & Hashimoto, T. Neural Schrödinger equation: physical law as deep neural network. Nikkhah, V. et al. Inverse-designed low-index-contrast structures on a silicon photonics platform for vector–matrix multiplication. & Feng, L. Lithography-free reconfigurable integrated photonic processor. Molesky, S. et al. Inverse design in nanophotonics. Psaltis, D., Brady, D., Gu, X.-G. & Lin, S. Holography in artificial neural networks. & Psaltis, D. Holographic interconnections in photorefractive waveguides. Delaney, M. et al. Nonvolatile programmable silicon photonics using an ultralow-loss Sb2Se3 phase change material. Wu, C. et al. Freeform direct-write and rewritable photonic integrated circuits in phase-change thin films. Delaney, M., Zeimpekis, I., Lawson, D., Hewak, D. W. & Muskens, O. L. A new family of ultralow loss reversible phase-change materials for photonic integrated circuits: Sb2S3 and Sb2Se3. Chiou, P. Y., Ohta, A. T. & Wu, M. C. Massively parallel manipulation of single cells and microparticles using optical images. & Wheeler, K. Acoustic characteristics of American english vowels. The MNIST Database of Handwritten Digits (New York Univ., accessed 07 January 2025); http://yann.lecun.com/exdb/mnist/ Deep physical neural networks trained with backpropagation. Programmable integrated photonic coherent matrix: principle, configuring, and applications. Clements, W. R., Humphreys, P. C., Metcalf, B. J., Kolthammer, W. S. & Walmsley, I. A. Optimal design for universal multiport interferometers. Hamerly, R., Basani, J. R., Sludds, A., Vadlamani, S. K, and Englund, D. To- ward the information-theoretic limit of programmable photonics. Li, S. & Hsu, C. W. Thickness bound for nonlocal wide-field-of-view metalenses. Gu, J. et al. M3icro: Ma- chine learning-enabled compact photonic tensor core based on programmable multi-operand multimode interference. Demonstration of scalable microring weight bank control for large-scale photonic integrated circuits. Zhang, H. et al. An optical neural chip for implementing complex-valued neural network. Pai, S. et al. Experimentally realized in situ backpropagation for deep learning in photonic neural networks. Inference in artificial intelligence with deep optics and photonics. & Engheta, N. Inverse-designed metastructures that solve equations. Roques-Carmes, C. et al. Heuristic recurrent algorithms for photonic ising machines. Hsu, C. W., Zhen, B., Stone, A. D., Joannopoulos, J. D. & Soljačić, M. Bound states in the continuum. Price, H. et al. Roadmap on topological photonics. Marcucci, G., Pierangeli, D. & Conti, C. Theory of neuromorphic computing by waves: machine learning by rogue waves, dispersive shocks, and solitons. Teğin, U., Yıldırım, M., Oğuz, İ., Moser, C. & Psaltis, D. Scalable optical learning operator. Bender, N. et al. Depth-targeted energy delivery deep inside scattering media. Phase and amplitude pulse shaping with two-dimensional phase-only spatial light modulators. Onodera, T. & Stein, M. Data repository of ‘Arbitrary control over multimode wave propagation for machine learning'. We acknowledge the US Air Force Office of Scientific Research for funding under award number FA9550-22-1-0378 and the National Science Foundation for funding under award number CCF-1918549. We thank NTT Research for their financial and technical support. This study was performed in part at the Cornell NanoScale Facility, a member of the National Nanotechnology Coordinated Infrastructure, which is supported by the National Science Foundation (grant number NNCI-2025233). acknowledges financial support from The David and Lucile Packard Foundation Fellowship. We acknowledge helpful discussions with C. Alpha, N. Bender, J. Clark, A. D'Addario, N. Flemens, J. Grazul, R. Hamerly, D. Heydari, P. Infante, O. Jaramillo, V. Kremenetski, M. Krenn, K. Li, G. McMurdy, R. Panepucci, C. Poitras, S. Prabhu, A. Windsor, F. Wu and Y. Zhao. Present address: Department of Electrical and Computer Engineering, Boston University, Boston, MA, USA Present address: Department of Electrical Engineering and Computer Science, University of California, Irvine, CA, USA Present address: Department of Applied Physics, Yale University, New Haven, CT, USA These authors contributed equally: Tatsuhiro Onodera, Martin M. Stein. School of Applied and Engineering Physics, Cornell University, Ithaca, NY, USA Tatsuhiro Onodera, Martin M. Stein, Benjamin A. Ash, Mandar M. Sohoni, Melissa Bosch, Ryotatsu Yanagimoto, Tianyu Wang, Gennady Shvets, Maxim R. Shcherbakov, Logan G. Wright & Peter L. McMahon Tatsuhiro Onodera, Martin M. Stein, Ryotatsu Yanagimoto, Marc Jankowski, Timothy P. McKenna & Logan G. Wright Kavli Institute at Cornell for Nanoscale Science, Cornell University, Ithaca, NY, USA Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar designed the devices and experiments. designed and built the imaging set-up to program the refractive index patterns. wrote the paper with input from all authors. Correspondence to Martin M. Stein or Peter L. McMahon. The other authors declare no competing interests. Nature Physics thanks Daniel Brunner and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Devices were fabricated from thin-film lithium niobate on insulator wafers consisting of a heavily doped p-type silicon substrate with a resistivity of 0.01 to 0.05Ω ⋅ cm, a 2 μm PECVD silicon dioxide layer, and 700 nm Z-cut MgO-doped lithium niobate bonded by ion-cutting. After dicing, we deposited 1 μm of SiO2 cladding and 4 μm of silicon-rich SiN (photoconductive layer) by PECVD with alternating plasma pulses to minimize stress. Electrodes were formed by evaporating 10 nm Ti and 5 nm Au, with tape masking to prevent breakdown at chip edges. Waveguide facets were polished using progressively finer abrasives to reduce coupling loss. The setup consists of five key units: (1) a beamshaper to generate spatially varying one-dimensional optical inputs to the 2D-programmable waveguide, (2) a digital micromirror projector to impose a programmable illumination pattern controlling the refractive-index distribution, (3) butt-coupling optics to couple light in and out of the device, (4) (not shown) a high-voltage source applying oscillating fields across the electrode and conductive silicon wafer, and (5) an infrared camera to record the output intensity profile. The beamshaper provides full amplitude and phase control of input fields, while the projector enables spatial refractive-index modulation with continuous pixel-values via pulse-width modulation. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Arbitrary control over multimode wave propagation for machine learning. Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. (2025)Cite this article Dendritic cells (DCs) are professional antigen-presenting cells. While plasmacytoid DCs (pDCs) are poor antigen-presenting cells at steady state, myeloid DCs (mDCs), which include DC1s, DC2s and DC3s, are specialized in T cell priming. To generate unbiased human DC atlases, we integrated DCs from 13 tumor tissues across 40 datasets to create a pDC + mDC-VERSE (DC-VERSE) and an mDC-VERSE single-cell RNA-sequencing compendium. We characterized DC subsets and ‘states' across these tissues. Most studied tumors contained CD207+ DCs, a subset of CD1c+ DCs, whose expansion inversely correlated with tumor CD8+ resident memory T cells, T cell clonality and the survival of patients treated with immune checkpoint inhibitors. Similarly to CCR7+ mDCs (a common state of DC1s, DC2s and DC3s), we found that CD207+ DCs were a common state of DC2s and DC3s. Spatially resolved single-cell transcriptomic and immunohistofluorescence analyses of human carcinomas demonstrated that lymphocytes and most DCs were enriched within the tumor stroma, while CD207+ DCs were mostly embedded within tumor nests. These DC-VERSEs provide a robust resource available to the scientific community on DCs in health and pathology. This is a preview of subscription content, access via your institution Access Nature and 54 other Nature Portfolio journals Get Nature+, our best-value online-access subscription cancel any time Subscribe to this journal Receive 12 print issues and online access $259.00 per year only $21.58 per issue Buy this article Prices may be subject to local taxes which are calculated during checkout No new data were generated; all analyzed datasets are public (Supplementary Table 1). The DC-VERSE and mDC-VERSE are available for download at https://github.com/gustaveroussy/FG-Lab. The DC-VERSE and the mDC-VERSE code can be found at https://github.com/gustaveroussy/FG-Lab. Steinman, R. M. & Cohn, Z. Identification of a novel cell type in peripheral lymphoid organs of mice: I. Morphology, quantitation, tissue distribution. Nussenzweig, M. C., Steinman, R. M., Gutchinov, B. A. Dendritic cells are accessory cells for the development of anti-trinitrophenyl cytotoxic T lymphocytes. Steinman, R. M. & Witmer, M. D. Lymphoid dendritic cells are potent stimulators of the primary mixed leukocyte reaction in mice. Steinman, R. M., Gutchinov, B., Witmer, M. D. & Nussenzweig, M. C. Dendritic cells are the principal stimulators of the primary mixed leukocyte reaction in mice. Guilliams, M. et al. Dendritic cells, monocytes and macrophages: a unified nomenclature based on ontogeny. Lee, J. et al. Restricted dendritic cell and monocyte progenitors in human cord blood and bone marrow. Liu, K. et al. In vivo analysis of dendritic cell development and homeostasis. Naik, S. H. et al. Development of plasmacytoid and conventional dendritic cell subtypes from single precursor cells derived in vitro and in vivo. Bourdely, P. et al. Transcriptional and functional analysis of CD1c+ human dendritic cells identifies a CD163+ subset priming CD8+CD103+ T cells. Cytlak, U. et al. Differential IRF8 transcription factor requirement defines two pathways of dendritic cell development in humans. et al. Single-cell analysis of human mononuclear phagocytes reveals subset-defining markers and identifies circulating inflammatory dendritic cells. See, P. et al. Mapping the human DC lineage through the integration of high-dimensional techniques. Villani, A.-C. et al. Single-cell RNA-seq reveals new types of human blood dendritic cells, monocytes, and progenitors. Ginhoux, F., Guilliams, M. & Merad, M. Expanding dendritic cell nomenclature in the single-cell era. Di Pilato, M. et al. CXCR6 positions cytotoxic T cells to receive critical survival signals in the tumor microenvironment. Lee, C. Y. C. et al. Tumour-retained activated CCR7+ dendritic cells are heterogeneous and regulate local anti-tumour cytolytic activity. Magen, A. et al. Intratumoral dendritic cell–CD4+ T helper cell niches enable CD8+ T cell differentiation following PD-1 blockade in hepatocellular carcinoma. Maier, B. et al. A conserved dendritic-cell regulatory program limits antitumour immunity. Moon, C. Y. et al. Dendritic cell maturation in cancer. Zheng, C. et al. Landscape of infiltrating T cells in liver cancer revealed by single-cell sequencing. Bell, D. et al. In breast carcinoma tissue, immature dendritic cells reside within the tumor, whereas mature dendritic cells are located in peritumoral areas. Franken, A. et al. CD4+ T cell activation distinguishes response to anti-PD-L1 + anti-CTLA4 therapy from anti-PD-L1 monotherapy. Leader, A. M. et al. Single-cell analysis of human non-small cell lung cancer lesions refines tumor classification and patient stratification. Zhang, Y. et al. Single-cell analyses reveal key immune cell subsets associated with response to PD-L1 blockade in triple-negative breast cancer. Mulder, K. et al. Cross-tissue single-cell landscape of human monocytes and macrophages in health and disease. Levine, J. H. et al. Data-driven phenotypic dissection of AML reveals progenitor-like cells that correlate with prognosis. Narasimhan, H. et al. RORγt-expressing dendritic cells are functionally versatile and evolutionarily conserved antigen-presenting cells. Rodrigues, P. F. et al. Rorγt-positive dendritic cells are required for the induction of peripheral regulatory T cells in response to oral antigens. Gao, Y. et al. Single-cell analysis reveals the heterogeneity of monocyte-derived and peripheral type-2 conventional dendritic cells. Brown, C. C. et al. Transcriptional basis of mouse and human dendritic cell heterogeneity. Cheng, S. et al. A pan-cancer single-cell transcriptional atlas of tumor infiltrating myeloid cells. Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cillo, A. R. et al. Immune landscape of viral- and carcinogen-driven head and neck cancer. Gerhard, G. M., Bill, R., Messemaker, M., Klein, A. M. & Pittet, M. J. Tumor-infiltrating dendritic cell states are conserved across solid human cancers. Zilionis, R. et al. Single-cell transcriptomics of human and mouse lung cancers reveals conserved myeloid populations across individuals and species. Qian, J. et al. A pan-cancer blueprint of the heterogeneous tumor microenvironment revealed by single-cell profiling. Zhang, Q. et al. Landscape and dynamics of single immune cells in hepatocellular carcinoma. Kvedaraite, E. & Ginhoux, F. Human dendritic cells in cancer. Merad, M., Sathe, P., Helft, J., Miller, J. & Mortha, A. The dendritic cell lineage: ontogeny and function of dendritic cells and their subsets in the steady state and the inflamed setting. Hao, Y. et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Lamb, J. et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease. Kvedaraite, E. et al. Notch-dependent cooperativity between myeloid lineages promotes Langerhans cell histiocytosis pathology. Bigley, V. et al. Langerin-expressing dendritic cells in human tissues are related to CD1c+ dendritic cells and distinct from Langerhans cells and CD141high XCR1+ dendritic cells. Wu, S. Z. et al. A single-cell and spatially resolved atlas of human breast cancers. Bassez, A. et al. A single-cell map of intratumoral changes during anti-PD1 treatment of patients with breast cancer. Ramos, R. N. et al. Tissue-resident FOLR2+ macrophages associate with CD8+ T cell infiltration in human breast cancer. Lánczky, A. & Győrffy, B. Web-based survival analysis tool tailored for medical research (KMplot): development and implementation. Duong, E. et al. Type I interferon activates MHC class I-dressed CD11b+ conventional dendritic cells to promote protective anti-tumor CD8+ T cell immunity. Garris, C. S. et al. Successful anti-PD-1 cancer immunotherapy requires T cell–dendritic cell crosstalk involving the cytokines IFN-γ and IL-12. Gabrilovich, D. Mechanisms and functional significance of tumour-induced dendritic-cell defects. Caetano, M. S. et al. IL6 blockade reprograms the lung tumor microenvironment to limit the development and progression of K-ras-mutant lung cancer. Cancer Res. Liu, Z. et al. Duality of interactions between TGF-β and TNF-α during tumor formation. Park, S. L., Gebhardt, T. & Mackay, L. K. Tissue-resident memory T cells in cancer immunosurveillance. Reading, J. L. et al. The function and dysfunction of memory CD8+ T cells in tumor immunity. Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. McInnes, L., Healy, J. & Melville, J. UMAP: uniform manifold approximation and projection for dimension reduction. Preprint at arXiv https://doi.org/10.48550/arXiv.1802.03426 (2018). MacParland, S. A. et al. Single cell RNA sequencing of human liver reveals distinct intrahepatic macrophage populations. Aizarani, N. et al. A human liver cell atlas reveals heterogeneity and epithelial progenitors. Ramachandran, P. et al. Resolving the fibrotic niche of human liver cirrhosis at single-cell level. Sharma, A. et al. Onco-fetal reprogramming of endothelial cells drives immunosuppressive macrophages in hepatocellular carcinoma. Tang-Huau, T.-L. et al. Human in vivo-generated monocyte-derived dendritic cells and macrophages cross-present antigens through a vacuolar pathway. Azizi, E. et al. Single-cell map of diverse immune phenotypes in the breast tumor microenvironment. Kim, D. et al. Targeted therapy guided by single-cell transcriptomic analysis in drug-induced hypersensitivity syndrome: a case report. Nakamizo, S. et al. Single-cell analysis of human skin identifies CD14+ type 3 dendritic cells co-producing IL1B and IL23A in psoriasis. Cheng, J. B. et al. Transcriptional programming of normal and inflamed human epidermis at single-cell resolution. He, H. et al. Mild atopic dermatitis lacks systemic inflammation and shows reduced nonlesional skin abnormalities. Xue, D., Tabib, T., Morse, C. & Lafyatis, R. Single-cell RNA sequencing reveals different subsets of macrophage and dendritic cells in human skin. Smillie, C. S. et al. Intra- and inter-cellular rewiring of the human colon during ulcerative colitis. Zhang, L. et al. Single-cell analyses inform mechanisms of myeloid-targeted therapies in colon cancer. James, K. R. et al. Distinct microbial and immune niches of the human colon. et al. Lineage-dependent gene expression programs influence the immune landscape of colorectal cancer. Reyfman, P. A. et al. Single-cell transcriptomic analysis of human lung provides insights into the pathobiology of pulmonary fibrosis. Vieira Braga, F. A. et al. A cellular census of human lungs identifies novel cell states in health and in asthma. Lambrechts, D. et al. Phenotype molding of stromal cells in the lung tumor microenvironment. Kim, N. et al. Single-cell RNA sequencing demonstrates the molecular and cellular reprogramming of metastatic lung adenocarcinoma. Arazi, A. et al. The immune cell landscape in kidneys of patients with lupus nephritis. Stewart, B. J. et al. Spatiotemporal immune zonation of the human kidney. Peng, J. et al. Single-cell RNA-seq highlights intra-tumoral heterogeneity and malignant progression in pancreatic ductal adenocarcinoma. Becht, E. et al. Dimensionality reduction for visualizing single-cell data using UMAP. Stuart, T. et al. Comprehensive integration of single-cell data. Van de Sande, B. et al. A scalable SCENIC workflow for single-cell gene regulatory network analysis. Stringer, C., Wang, T., Michaelos, M. & Pachitariu, M. Cellpose: a generalist algorithm for cellular segmentation. Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. Mackay, L. K. et al. Hobit and Blimp1 instruct a universal transcriptional program of tissue residency in lymphocytes. Download references We thank L. Robinson of Insight Editing London for the critical review and editing of the manuscript. We thank Y. Velut for providing the immunohistofluorescence images, as well as the Cell Imaging and Flow Cytometry Platform (CHIC) of the Centre de Recherche des Cordeliers for its help with this study. We thank the Foundation MSD Avenir (https://www.msdavenir.fr/) for its financial contribution to this project. We thank the Marie Lannelongue Hospital and its biobank for their valuable collaboration and support of this study. This work was supported by INSERM, Sorbonne Université, Université de Paris, Ligue Contre le Cancer (Equipe Labellisée), the CARPEM (Cancer Research for Personalized Medicine) program of the Sites Intégrés de Recherche sur le Cancer (SIRIC), and LabEx Immuno-Oncology. F.G. is an EMBO YIP awardee and is supported by Singapore Immunology Network (SIgN) core funding as well as a Singapore National Research Foundation Senior Investigatorship (NRFI) NRF2016NRF-NRFI001-02 and the Foundation Gustave Roussy. C.-A.D. is an INSERM researcher supported by INSERM. was supported by the European Union's Horizon 2020 research and innovation program under grant agreement no. 825410 (ONCOBIOME project), ANR RHU5 ‘ANR-21-5 RHUS-0017' IMMUNOLIFE, MAdCAM INCA_ 16698 and ERC advanced, funded by the European Research Council (ERC) under grant agreement number 101052444, the ANR-23-RHUS-0010 (LUCA-pi), the European Union's Horizon 2020 research and innovation program no. 964590 (project acronym: IHMCSA, project title: International Human Microbiome Coordination and Support Action), the European Union's Horizon Europe research and innovation program under grant agreement no. 101095604 (project acronym: PREVALUNG EU, project title: Personalized lung cancer risk assessment leading to stratified interception), as well as by the SEERAVE Foundation. Other grant supports include Ligue Contre le Cancer and the SIGN'IT ARC Foundation (MICROBIONT-PREDICT, 2021). Present address: INSERM U955, Institut Mondor de Recherche Biomédicale (IMRB), Université Paris-Est Créteil, Créteil, France These authors contributed equally: Kevin Mulder, Margaux Gardet. These authors jointly supervised this work: Florent Ginhoux, Charles-Antoine Dutertre. INSERM U1015, Gustave Roussy, Paris-Saclay University, Villejuif, France Kevin Mulder, Margaux Gardet, Wan Ting Kong, Amit Ashok Patel, Grégoire Gessain, Carlos de la Calle-Fabregat, Elisa Poupaud, Ahmed-Amine Anzali, Garett Dunsmore, Anne-Gaëlle Goubet, Laurence Zitvogel, Florent Ginhoux & Charles-Antoine Dutertre Centre de Recherche des Cordeliers, INSERM, Sorbonne Université, USPC Université Paris Cité, Equipe Labellisée Ligue Nationale Contre le Cancer, Paris, France Anne Calvez, Antoine Bougouin, Guilhem Pupier, Catherine Sautès-Fridman & Wolf H. Fridman Faculté de Santé, Université Paris Cité, Paris, France Immunobiology Laboratory, The Francis Crick Institute, London, UK INSERM U981, PRISM Center, Gustave Roussy, Paris-Saclay University, Villejuif, France Laboratory of Mathematics and Computer Science (MICS), CentraleSupélec, Paris-Saclay University, Gif-sur-Yvette, France Vizgen, Cambridge, MA, USA Lizhe He, Timothy Wiggins, Jiang He & George Emanuel Broad Institute of MIT and Harvard, Cambridge, MA, USA Cancer Ecosystems Program, Garvan Institute of Medical Research, Darlinghurst, New South Wales, Australia School of Clinical Medicine, Faculty of Medicine & Health, UNSW Sydney, Sydney, New South Wales, Australia INSERM U1186, Gustave Roussy, Paris-Saclay University, Villejuif, France Institute of Systems Immunology, Hamburg Center for Translational Immunology (HCTI), University Medical Center Hamburg-Eppendorf, Hamburg, Germany Regine J. Plateforme de Bioinformatique, INSERM US23, CNRS UMS 3655, Université Paris-Saclay, Villejuif, France Program in Emerging Infectious Disease, Duke-NUS Medical School, Singapore, Singapore Pathology Department, Marie Lannelongue Center, Le Plessis Robinson, France Vincent Thomas de Montpreville Singapore Immunology Network (SIgN), A*STAR, Singapore, Singapore Shanghai Institute of Immunology, Shanghai Jiao Tong University School of Medicine, Shanghai, China Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Provision of human NSCLC samples: A.-G.G. Provision of human NSCLC FFPE blocks: V.T.d.M. Provision of human ovarian samples: C.d.l.C.-F. and J.M. Provision of breast cancer Visium spatial transcriptomic data: A.S. Generation, provision and segmentation of MERFISH data: L.H., T.W., J.H. Establishment of the publicly available online DC-VERSE and mDC-VERSE: K.M. Writing of the manuscript: K.M., M.G., W.T.K., A.A.P., F.G. and C.-A.D. Online cellXgene VERSEs: M.D. Project supervision: F.G. and C.-A.D. Study conceptualization: F.G. and C.-A.D. Correspondence to Florent Ginhoux or Charles-Antoine Dutertre. F.G. and C.-A.D. are coinventors of a patent related to the findings described in this article. The other authors declare no competing interests. Nature Immunology thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editor: Ioana Staicu, in collaboration with the Nature Immunology team. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Violin plots displaying nFeature RNA, nCounts RNA, the percentage of mitochondrial genes (percent.mito), the percentage of ribosomal genes (percent.ribo), and the percentage of heat shock protein genes (percent.hs). Related to Fig. a, Phenograph clusters' (cl.) annotation of the mDC-VERSE. b, Heatmap showing the relative expression levels of Differentially Expressed Regulons (DERs) between phenograph clusters common to Lung (Maier) and Tonsil (Cillo) cancer datasets. c, Quality control metrics for each Phenograph cluster and meaning plot of nFeature_RNA. d, Annotation of cl. #15 on the mDC-VERSE. e-f, CITE-seq data (from Maier et al.) showing expression of signature T and B cell protein markers and DC2 and DC3 protein markers. g, Identification of DC2s and DC3s using CD5 and CD14 protein expression from CITE-seq data (from Maier et al.) within the DC2 + DC3 region of the mDC-VERSE. h, Meaning plots of DC2 and DC3 gene signatures from Dutertre et al. on the mDC-VERSE. i, DEG heatmap between mega-clusters of the mDC-VERSE. j, Mean expression of the moDC signature from Gao et al. overlayed onto the MNP-VERSE (from Mulder et al.) and onto the mDC-VERSE UMAP spaces. k, moDC signature score for each cell of the different mDC-VERSE Phenograph clusters. l,m, Composition of DC mega clusters across juxta-tumoral “healthy” tissues. n, Annotation of cl.3 & cl.4 (corresponding to cDC2As_Brown) from Brown et al. on the mDC-VERSE. Related to Fig. a-b, Projection of a, all mDC subsets and b, DC2 + DC3 populations defined by Cheng et al.'s metadata on the mDC-VERSE using multimodal reference mapping. c, Quality control metrics of predicted mDC-VERSE Phenograph clusters from Cheng et al. data projected by multimodal reference mapping. d, Mean expression of the top 50 genes of mega-clusters from the mDC-VERSE mapped onto the UMAP from Cheng et al. e, Mapping of cDC2_CD1A cells from Cheng et al., enriched in CD207 or LTB DC signatures (from the mDC-VERSE) onto the UMAP from Cheng et al. f, Meaning plots of the mean gene signatures of DC2 + DC3 populations from Cheng et al. shown on the mDC-VERSE. Related to Fig. a, Density plots of global colon, liver and lung datasets highlighting changes in DC1s, CCR7 mDCs, CD207 DCs, Prolif. DCs, ISG DCs and LTB DCs between juxta-tumoral and tumoral tissues. b, Percentage of CCR7 mDCs and Prolif. DCs, and DC3/DC2 ratio in datasets which had analysed juxta-tumoral tissue, tumour periphery and tumour core. c,d, Percentage of mDC-VERSE c, phenograph clusters and d, mega clusters in all integrated and query datasets (Obtained through multimodal reference mapping and annotated with cross symbol) between matched juxta-tumoral and cancer tissues. See Supplementary Table 1 for the specified tumour types. P-values were calculated using a Wilcoxon non-parametric paired test. Related to Fig. a, Gating strategy from singlets, live, CD45+ cells and projection of each gated population onto the Live_CD45+_UMAP space. b, MNP extracted from the Live_CD45+_UMAP were analysed by UMAP to generate the MNP_UMAP, whose annotation is confirmed by protein expression. c, Gating of pDCs and pre-DCs within CD123+ DCs defined in Fig. d, RNA expression of CD207 and CD1A and protein expression of CD103 overlaid on the mDC-VERSE. e, Fold increase of CD207+ DCs in tumour versus matched juxta-tumoral tissue. Related to Fig. a, Mapping of cMAP scores from Fig. 5b on the mDC-VERSE. b, Overlay of DC “states” identified in Fig. 5c onto the MNP_UMAP space. c, Expression of CADM1 and CD141 by CCR7+ mDCs, CD103+ DCs and CD207+ DCs. d, Gating and phenotype of CD103+ “LTB” and CD1a+CD207+ DCs. e, Expression of CD45, CD1a, CD1c, HLA-DR, HLA-DP, CD88 and CD3/CD16/CD19/CD20 versus CD207 by total live cells (including CD45− non-immune cells) from a NSCLC tumour. f, Mean fluorescence intensity (MFI) of markers expressed by populations of DCs defined in panel d. g, Percentage of DC “states” identified in (Fig. 5c) among total CD45+ cells in matched juxta-tumoral tissue versus tumour. h, Gene set enrichment analysis (GSEA) of the CD207 DC signature comparing bulk RNAseq of DC3s at day 3 cultured with GM-CSF + TGF-β with or without OP-9-D4 cells from Kvedaraite et al., 2022. (i) Gating strategy from singlets for the sorting of CD207+ DCs in ovarian cancer. P-values were calculated using a Wilcoxon non-parametric paired test, two-tailed. Related to Fig. a, Visium spatial transcriptomic profiling of 3 TNBC and 2 ER breast cancer patients from Wu et al. For each patient, the left panel shows the CD207 DC signature score, middle panel shows tissue niches, and the right panel shows haematoxylin and eosin (H&E) staining. b,) Enrichment score of CD207 DC signature across different tissue niches identified in a. c, Meaning plots of EPCAM and PTPRC expression visualised on the UMAP generated with all cells from the Merscope data of the breast cancer patient. Immune cells were extracted, and different cell populations were annotated based on a curated list of genes. mDCs were then extracted to generate an mDC UMAP that identified mDC populations. d, Meaning plots of representative genes used to define the immune populations identified in the Immune cells' UMAP from panel c. e,f, Merfish analysis of breast cancer and lung cancer cross-sections. e, Visualisation of the expression of DC population-defining transcripts in the segmented Merfish spatial data. f, Spatial distribution of tumour cells (grey) and immune populations within the breast and lung tumour cross-sections analysed by Merfish. g, Single fluorescent images for CD207 (green), CD3 (red), CD8 (yellow) and CD20 (cyan) of the IHF data shown in Fig. Related to Fig. a, Gating strategy for identifying mDC populations in the ICS experiment (see Fig. b, Percentage of positive cells for co-stimulatory factors in, (Left) CD1c+ DC2/3 CD207 +/−, (Middle) CADM1+ DC1 CCR7+/−,(Right) CD1c+DC2/3 CCR7+/−. P-values were calculated using a Wilcoxon non-parametric paired test, two-tailed. Related to Fig. a, Percentage of predicted phenograph clusters from query dataset (Bassez et al.) projected using multimodal reference mapping onto the mDC-VERSE. b, Percentage of predicted DC2 and DC3 mega-clusters by multimodal reference mapping of query data from breast cancer patients categorised by T-cell clonality and treatment status (anti-PD-1 therapeutic monoclonal antibody = Immune Checkpoint Blockade = ICB). c, Percentage of CD207 DCs, CCR7 mDC, ISG DC and DC1 between patients with non-expanded and expanded T-cell clonality in the Bassez et al. data. d, Upper panel shows the correlation between the frequencies among CD45+ cells of DC populations (DC1s, ISG DCs, DC2s, DC3s and CCR7 mDCs) and CD8 TRMs in lung tumours within the Leader et al. scRNAseq data. Lower panel shows the frequencies among PTPRC(CD45)-expressing immune cells of the same DC populations split by CD8 TRMshi and CD8 TRMslo. e, Correlation between the frequencies of CD207+ DCs and CD4+ T-cells from flow cytometry analysis of 8 NSCLC patients. f, Correlation map of DC population signatures (defined in the mDC-VERSE) and of other signatures obtained from Ramos et al. in the BRCA (Breast) and the LUAD (Lung) adenocarcinoma TCGA datasets. g, Kaplan-Meier plots of the overall survival (OS) of patients with different cancers whose tumour was sampled and analysed by bulk RNAseq prior to immune checkpoint blockade (ICB) treatment. Patients were separated based on high or low expression of genes specifically expressed by total DC2s + DC3s, by CD207 DCs, by DC1s or by CCR7 mDCs. Correlations were evaluated using the Pearson correlation (r) with two-tailed p values. P-values were calculated using a Wilcoxon non-parametric paired test, two-tailed. Related to Fig. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Reprints and permissions Mulder, K., Gardet, M., Kong, W.T. et al. DC subsets and states unraveled across human juxtatumoral and malignant tissues. Version of record: 08 December 2025 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative © 2025 Springer Nature Limited Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Synthesizing perceivable artificial neural inputs independent of typical sensory channels remains a fundamental challenge in developing next-generation brain−machine interfaces. Establishing a minimally invasive, wirelessly effective and miniaturized platform with long-term stability is crucial for creating research methods and clinically meaningful biointerfaces capable of mediating artificial perceptual feedback. Here we demonstrate a miniaturized, fully implantable transcranial optogenetic neural stimulator designed to generate artificial perceptions by patterning large cortical ensembles wirelessly in real time. Experimentally validated numerical simulations characterized light and heat propagation, whereas neuronal responses were assessed by in vivo electrophysiology and molecular methods. Cue discrimination during operant learning demonstrated the wireless genesis of artificial percepts sensed by mice, where spatial distance across large cortical networks and sequential order-based analyses of discrimination predicted performance. These conceptual and technical advances expand understanding of artificially patterned neural activity and its perception by the brain to guide the evolution of next-generation all-optical brain−machine communication. This is a preview of subscription content, access via your institution Get Nature+, our best-value online-access subscription cancel any time Subscribe to this journal Receive 12 print issues and online access Buy this article Prices may be subject to local taxes which are calculated during checkout Raw data generated during the present study are available from the corresponding authors upon reasonable request. The analyzed data are available at https://doi.org/10.5281/zenodo.14880024 (ref. Data supporting the findings of this study are included within this paper and its Supplementary Information files. Source data are provided with this paper. All computer code and customized software generated during and/or used in the present study are available at https://doi.org/10.5281/zenodo.14880024 (ref. & Nicolelis, M. A. L. Brain–machine interfaces: past, present and future. Sense of agency for intracortical brain–machine interfaces. Tang, X., Shen, H., Zhao, S., Li, N. & Liu, J. Active tactile exploration using a brain–machine–brain interface. Flesher, S. N. et al. A brain-computer interface that evokes tactile sensations improves robotic arm control. & Constandinou, T. G. Implantable brain machine interfaces: first-in-human studies, technology challenges and trends. Flexible near-field wireless optoelectronics as subdermal implants for broad applications in optogenetics. Yang, Y. et al. Wireless multilateral devices for optogenetic studies of individual and social behaviors. Ausra, J. et al. Wireless, battery-free, subdermally implantable platforms for transcranial and long-range optogenetics in freely moving animals. Preparation and use of wireless reprogrammable multilateral optogenetic devices for behavioral neuroscience. Wu, Y. et al. Wireless multi-lateral optofluidic microsystems for real-time programmable optogenetics and photopharmacology. Marshel, J. H. et al. Cortical layer–specific critical dynamics triggering perception. Targeted activation of hippocampal place cells drives memory-guided spatial behavior. Precise holographic manipulation of olfactory circuits reveals coding features determining perceptual detection. Carrillo-Reid, L., Han, S., Yang, W., Akrouh, A. & Yuste, R. Controlling visually guided behavior by holographic recalling of cortical ensembles. The logic of recurrent circuits in the primary visual cortex. Adesnik, H. & Abdeladim, L. Probing neural codes with two-photon holographic optogenetics. Prominent in vivo influence of single interneurons in the developing barrel cortex. Chong, E. et al. Manipulating synthetic optogenetic odors reveals the coding logic of olfactory perception. Pinto, L. & Dan, Y. Cell-type-specific activity in prefrontal cortex during goal-directed behavior. Pinto, L. et al. Task-dependent changes in the large-scale dynamics and necessity of cortical regions. Pinto, L., Tank, D. W. & Brody, C. D. Multiple timescales of sensory-evidence accumulation across the dorsal cortex. B. Striatal circuits for reward learning and decision-making. Ouyang, W. et al. A wireless and battery-less implant for multimodal closed-loop neuromodulation in small animals. Stujenske, J. M., Spellman, T. & Gordon, J. Modeling the spatiotemporal dynamics of light and heat propagation for in vivo optogenetics. Owen, S. F., Liu, M. H. & Kreitzer, A. C. Thermal constraints on in vivo optogenetic manipulations. Wang, Q. et al. The Allen Mouse Brain Common Coordinate Framework: a 3D reference atlas. Independent optical excitation of distinct neural populations. & Shenoy, K. V. ERAASR: an algorithm for removing electrical stimulation artifacts from multielectrode array recordings. Nakayama, H., Gerkin, R. C. & Rinberg, D. A behavioral paradigm for measuring perceptual distances in mice. Harris, K. D. & Mrsic-Flogel, T. D. Cortical connectivity and sensory coding. Variation in spatial dependencies across the cortical mantle discriminates the functional behaviour of primary and association cortex. Paolino, A. et al. Non-uniform temporal scaling of developmental processes in the mammalian cortex. Levi, A. J., Yates, J. L., Huk, A. C. & Katz, L. N. Strategic and dynamic temporal weighting for perceptual decisions in humans and macaques. Hyafil, A. et al. Temporal integration is a robust feature of perceptual decisions. Resulaj, A., Ruediger, S., Olsen, S. R. & Scanziani, M. First spikes in visual cortex enable perceptual discrimination. Nienborg, H. & Cumming, B. G. Decision-related activity in sensory neurons reflects more than a neuron's causal effect. Weber, A. I. et al. Spatial and temporal codes mediate the tactile perception of natural textures. Gavornik, J. P. & Bear, M. F. Learned spatiotemporal sequence recognition and prediction in primary visual cortex. Baker, C. A., Elyada, Y. M., Parra, A. & Bolton, M. M. Cellular resolution circuit mapping with temporal-focused excitation of soma-targeted channelrhodopsin. Daigle, T. L. et al. A suite of transgenic driver and reporter mouse lines with enhanced brain-cell-type targeting and functionality. High-performance microbial opsins for spatially and temporally precise perturbations of large neuronal networks. & Deisseroth, K. Integration of optogenetics with complementary methodologies in systems neuroscience. Ben-Simon, Y. et al. A suite of enhancer AAVs and transgenic mouse lines for genetic access to cortical cell types. Shin, H. et al. Transcranial optogenetic brain modulator for precise bimodal neuromodulation in multiple brain regions. McCue, A. C. & Kuhlman, B. Design and engineering of light-sensitive protein switches. Selective attention and the organization of visual information. Lemus, L., Hernández, A., Luna, R., Zainos, A. & Romo, R. Do sensory cortices process more than one sensory modality during perceptual judgments? O'Riordan, M. & Passetti, F. Discrimination in autism within different sensory modalities. Effect of same- and different-modality spatial cues on auditory and visual target identification. Karlin, L. & Mortimer, R. G. Effect of verbal, visual, and auditory augmenting cues on learning a complex motor skill. Niyo, G., Almofeez, L. I., Erwin, A. A computational study of how an α- to γ-motoneurone collateral can mitigate velocity-dependent stretch reflexes during voluntary movement. Edge computing in nature: minimal pre-processing of multi-muscle ensembles of spindle signals improves discriminability of limb movements. & Sathian, K. Temporal cues contribute to tactile perception of roughness. Hirsh, I. J. Auditory perception of temporal order. Greenspon, C. M., Shelchkova, N. D., Hobbs, T. G., Bensmaia, S. J. & Gaunt, R. A. Intracortical microstimulation of human somatosensory cortex induces natural perceptual biases. Partial recovery of visual function in a blind patient after optogenetic therapy. Drew, L. Restoring vision with optogenetics. Taal, A. J. et al. Optogenetic stimulation probes with single-neuron resolution based on organic LEDs monolithically integrated on CMOS. Wang, Y., Garg, R., Cohen-Karni, D. & Cohen-Karni, T. Neural modulation with photothermally active nanomaterials. & Wehr, M. AUTOPILOT: automating experiments with lots of Raspberry Pis. Preprint at bioRxiv https://doi.org/10.1101/807693 (2019). Lu, L., Leanza, S. & Zhao, R. R. Origami with rotational symmetry: a review on their mechanics and design. He, D., Malu, D. & Hu, Y. A Comprehensive review of indentation of gels and soft biological materials. Yan, P., Huang, H., Meloni, M., Li, B. Mechanical properties inside origami-inspired structures: an overview. Christensen, R. M. Review of the basic elastic mechanical properties and their realignment to establish ductile versus brittle failure behaviors. Huang, Y. et al. Microfluidic serpentine antennas with designed mechanical tunability. Analysis and management of thermal loads generated in vivo by miniaturized optoelectronic implantable devices. & Boas, D. A. Monte Carlo simulation of photon migration in 3D turbid media accelerated by graphics processing units. Zhang, H. et al. Analytical solutions for light propagation of LED. Deshpande, V. S. & McMeeking, R. M. Models for the interplay of mechanics, electrochemistry, thermodynamics, and kinetics in lithium-ion batteries. Zhao, W., Liu, L., Lan, X., Leng, J. & Liu, Y. Thermomechanical constitutive models of shape memory polymers and their composites. Kwon, K. et al. Wireless, soft electronics for rapid, multisensor measurements of hydration levels in healthy and diseased skin. Wu, M. et al. Attenuated dopamine signaling after aversive learning is restored by ketamine to rescue escape actions. Wu, M., Minkowicz, S., Dumrongprechachan, V., Hamilton, P. & Kozorovitskiy, Y. Ketamine rapidly enhances glutamate-evoked dendritic spinogenesis in medial prefrontal cortex through dopaminergic mechanisms. Xiao, L., Priest, M. F., Nasenbeny, J., Lu, T. & Kozorovitskiy, Y. Biased oxytocinergic modulation of midbrain dopamine systems. Schindelin, J. et al. Fiji: an open-source platform for biological-image analysis. Pachitariu, M., Sridhar, S., Pennington, J. & Stringer, C. Spike sorting with Kilosort4. Siegle, J. H. et al. Open Ephys: an open-source, plugin-based platform for multichannel electrophysiology. Pinto, L. et al. An accumulation-of-evidence task using visual pulses for mice navigating in virtual reality. Rodriguez, A. et al. ToxTrac: a fast and robust software for tracking organisms. Wu, M. et al. Dopamine pathways mediating affective state transitions after sleep loss. Wu, M. et al. Data and code for the article ‘Patterned wireless transcranial optogenetics generates artificial perception'. We thank L. Butler for mouse colony management and F. Valero-Cuevas for meaningful insights and discussions. This work made use of the NUFAB facility of Northwestern University's NUANCE Center, which has received support from the SHyNE Resource (National Science Foundation (NSF) ECCS-2025633), the International Institute for Nanotechnology and Northwestern's Materials Research Science and Engineering Center program (NSF DMR-2308691). MicroCT imaging work was performed at the Northwestern University Center for Advanced Molecular Imaging (RRID: SCR_021192), generously supported by National Cancer Institute Cancer Center Support Grant P30 CA060553 awarded to the Robert H. Lurie Comprehensive Cancer Center. Microscopy analyses using Leica SP8 were performed at the Biological Imaging Facility at Northwestern University (RRID: SCR_017767), generously supported by the Chemistry for Life Processes Institute, the Northwestern University Office for Research, the Department of Molecular Biosciences and the Rice Foundation. This work was funded by the Querrey-Simpson Institute for Bioelectronics (M.W., Y.Y., A.I.E., A.V.-G., Y.W., J.G., L.Z., J.L., M.K., J.K., Y.H. ); National Institute of Neurological Disorders and Stroke (NINDS)/BRAIN Initiative 1U01NS131406 (Y.K. ); National Institute of Mental Health (NIMH) R01MH117111 (Y.K. ); 2021 One Mind Nick LeDeit Rising Star Research Award (Y.K. ); Shaw Family Pioneer Award; Center for Reproductive Science, Feinberg School of Medicine (J.M.C. ); NIMH R00MH120047 (L.P.); Simons Foundation grant 872599SPI (L.P.); Alfred P. Sloan Foundation grant SP-2022-19027 (L.P.); North Carolina State University Start-up Fund 201473-02139 (A.V.-G.); 2T32MH067564 (J.Z. ); and the Christina Enroth-Cugell and David Cugell fellowship (M.W.). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript. These authors contributed equally: Mingzheng Wu, Yiyuan Yang, Jinglan Zhang, Andrew I. Efimov, Xiuyuan Li, Kaiqing Zhang. Querrey Simpson Institute for Bioelectronics, Northwestern University, Evanston, IL, USA Mingzheng Wu, Yiyuan Yang, Andrew I. Efimov, Yue Wang, Jianyu Gu, Glingna Wang, Minsung Kim, Liangsong Zeng, Jiaqi Liu, Minkyu Lee, Jiheon Kang, Joanna L. Ciatti, Kaila Ting, Stephen Cheng, Anthony Banks, Cameron H. Good, Abraham Vázquez-Guardado, Yonggang Huang & John A. Rogers Department of Neurobiology, Northwestern University, Evanston, IL, USA Mingzheng Wu, Jinglan Zhang, Kevin L. Bodkin, Lauren H. Yoon, Sara N. Freda, Lucas Pinto & Yevgenia Kozorovitskiy Department of Mechanical Engineering, Northwestern University, Evanston, IL, USA Yiyuan Yang, Liangsong Zeng, Yonggang Huang & John A. Rogers Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA Department of Biomedical Engineering, Northwestern University, Evanston, IL, USA Andrew I. Efimov, Yue Wang & John A. Rogers Department of Civil and Environmental Engineering, Northwestern University, Evanston, IL, USA Xiuyuan Li, Kaiqing Zhang, Haohui Zhang & Yonggang Huang State Key Laboratory of Mechanical System and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai, China Xiuyuan Li & Wenming Zhang State Key Laboratory of Structural Analysis, Optimization and CAE Software for Industrial Equipment, Department of Engineering Mechanics, Dalian University of Technology, Dalian, China Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, USA Mohammad Riahi & Abraham Vázquez-Guardado Center for Advanced Self-Powered Systems of Integrated Sensors and Technologies (ASSIST), North Carolina State University, Raleigh, NC, USA Mohammad Riahi & Abraham Vázquez-Guardado Pritzker School of Molecular Engineering, The University of Chicago, Chicago, IL, USA Computation and Neural Systems, California Institute of Technology, Pasadena, CA, USA Department of Materials Science and Engineering, Northwestern University, Evanston, IL, USA Joanna L. Ciatti, Yonggang Huang & John A. Rogers Department of Biomedical Engineering and the Institute of Materials Science, University of Connecticut, Storrs, CT, USA Center for Bio-Integrated Electronics, Northwestern University, Evanston, IL, USA Anthony Banks, Cameron H. Good & John A. Rogers Neurolux, Inc., Northfield, IL, USA Anthony Banks & Cameron H. Good Department of Neuroscience, Feinberg School of Medicine, Northwestern University, Chicago, IL, USA Julia M. Cox & Lucas Pinto Department of Neurobiology, The University of Chicago, Chicago, IL, USA Julia M. Cox & Lucas Pinto Chemistry of Life Processes Institute, Northwestern University, Evanston, IL, USA Department of Neurological Surgery, Feinberg School of Medicine, Northwestern University, Chicago, IL, USA Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar contributed equally to this work. Correspondence to Yiyuan Yang, Abraham Vázquez-Guardado, Yonggang Huang, Yevgenia Kozorovitskiy or John A. Rogers. are co-founders in a company, Neurolux, Inc., that offers related technology products to the neuroscience community. is employed by Neurolux, Inc. The other authors declare no competing interests. Nature Neuroscience thanks Luis Carrillo-Reid and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. (a) Left, schematic illustration of the circuit base of the electronic module of the device using three-layer flexible printed circuit board (fPCB). Right, images of the fPCB for the electronic module. (b) Left, schematic illustration of the assembly of electronic components on the fPCB of the electronic module using hot-air soldering. Right, images of the electronic module after hot-air soldering of electronic components. (c) Left, schematic illustration of the circuit base of the FOD using three-layer fPCB. Right, magnified view of the soldering site for the μ-ILED with solder paste. (d) Left, schematic illustration of the FOD after hot-air soldering of the μ-ILEDs. Right, magnified view of the soldered μ-ILED. (e) Left, schematic illustration of the serpentine traces of the device using a five-layer fPCB. Right, image of the serpentine traces on the fPCB. (f) Left, image of the serpentine traces after laser ablation from the fPCB substrate. Right, scanning electron microscopic (SEM) image of the serpentine traces from lateral view. The experiment was repeated independently on 3 serpentine traces. (g) Schematic illustration of the final assembly of the electronic module, serpentine traces, and FOD. (h) Image of the final device after assembly. (i) Schematic illustration of the mold used for final silicone (Ecoflex 00-30) coating after parylene-C encapsulation. (j) Photograph of the device during mold casting of silicone. (k) Image of a fully encapsulated device ready for in vivo experiments. (a) Left, geometric model of structures surrounding the serpentine interconnect. Right, calculation of the applied strain on the serpentine and equivalent strain on the Cu-based serpentine conductive traces. (b) Layered structure of the serpentine materials relative to the neutral mechanical plane. (c) Equivalent strain distribution on copper traces with 13% applied strain. (d) When a certain strain is applied to the serpentine interconnect, a certain copper unit on the trace shows the highest equivalent strain among all units. Summary graph showing this highest equivalent strain versus applied strains. Equivalent strain for copper plastic deformation equals 0.3%. (e) Geometric parameters that affect the stretchability of the serpentine interconnects. (f) Serpentine design examples with different gap widths and radius. (g) Contour map summarizing the stretchability as a function of normalized gap width and radius. Stretchability: maximal applied strain without causing plastic deformation on copper-based conductive traces. (h) Average elongation of the gap when the serpentine interconnect is stretched. (i) Elongation of the gap results in principal strain on the elastomer, causing potential encapsulation defects. (j) Design 1 & 2: prototypes, not guided by numerical modeling; Design 3, optimized layer structure with preliminary modifications of gap width and connecting regions; Design 4, optimized layer structure and geometry guided by numerical modeling. (k) Equivalent strain on the Cu-based conductive traces versus applied strain on the serpentine interconnect for the four designs. Dashed line: Equivalent strain threshold for plastic deformation on the copper traces. (l) Finite element analysis (FEA) of equivalent strain on the optimized serpentine interconnect under stretching (left) and bending (middle). (m) Schematic illustration of the benchtop validation of the modeling outcomes using cyclic stretching and bending tests of devices integrated on artificial skin immersed in saline. (n) Summary data showing normalized conductance loads on individual μ-ILEDs for 20k cycles in cyclic stretching and bending tests. n = 21 pairs of traces from 3 devices for design 1; n = 27 pairs of traces from 3 devices for design 2; n = 26 pairs of traces from 3 devices for design 3; n = 20 pairs of traces from 3 devices for design 4. Data are presented as mean ± s.e.m. (o) Summary data for stable cycles where the devices maintain constant resistance during stretching and bending. Dots represent individual devices. (p) Images showing devices in operation in the Bpod chamber, 32 days after implantation. (a) Schematic illustrations showing electronic circuits for wireless power harvesting and voltage regulation. (b) Capacitor bank for energy storage and discharge. (c) Near-field communication module for real-time programming. (d) Micro-controller circuit for parameter (order, frequency, duty cycle) control. (e) Digital-to-analog converters for intensity modulation. Essential components and input/output pins are labeled on the schematics. (a) Simulated magnetic-field-intensity distribution at the central plane of a behavioral cage (dimensions, 20 cm (length) × 14 cm (width)) with a double-loop antenna at heights of 3 cm and 6 cm. Scale bar: 10 cm. (b) Simulated magnetic-field-intensity distribution at different heights in the Bpod behavioral cage. (c) The total electrical power of the μ-ILED with respect to the primary antenna power. (e) The optical power of the μ-ILED with respect to the input current. (f) Schematic illustration (left) and image (right) of the experimental setup for measuring light attenuation through the skull and brain tissue. (g) Simulated (left) and measured (right) results of light attenuation through the skull. A 60 µm layer of skull was used in the numerical model. A piece of thinned skull was used for measurement. (h) Illumination volume and penetration depth as a function of the input irradiance of the red μ-ILED (628 nm). (i) Contour map plot showing total overlap volume from 4 μ-ILED co-activation for different input optical power and intensity thresholds for opsin variants. (a) Left, image of micro-fabricated thermistor for measuring temperature; middle, schematic illustration of the experimental setup for measuring heat accumulation in the structures surrounding the optical-neural interface; right, image of experimental setup, with wires connected to the thermistors to collect resistance value. (b) Calibration curves of thermistors used in this study for measuring the temperature increase on the surface of skull (left), on the surface of brain (middle), and below the µ-ILED (right). Each dot represents one resistance measurement at one temperature measurement. The best-fit line represents the linear regression between resistance and temperature. n = 1 representative thermistor at each location (skull, brain, or µ-ILED). (c) Simulated (Top) and measured (bottom) results of temperature increase below the µ-ILED for 3 s μ-ILED operation. The temperature increases in a 20 × 300 × 300 µm3 volume, 100 µm below the μ-ILED surface, was output from the numerical model. The thermistor was manually placed and adhered to the μ-ILED surface, followed by a complete procedure of μ-ILED array fabrication. Measured data are presented as mean ± s.e.m. Technical replicates are included to account for fluctuations in environmental temperature. (d) Left, simulated maximal temperature increases on the skull surface during μ-ILED operation with 10% duty cycle at varying frequencies. Right, same as left, but for varying duty cycles at 10 Hz. The finite element node with maximal temperature increase (max. node) was selected for plotting. (e) Same as (d), but for brain surface. (f) Left, measured temperature increases on the skull surface during μ-ILED operation with 10% duty cycle at varying frequencies. Right, same as left, but for varying duty cycles at 10 Hz. (g) Same as (f), but for brain surface. (h) Heatmap showing simulated heat production during a single μ-ILED operation in air at 100% duty cycle and 1.93 mW power for 10 s. Left, fPCB array patch without modification. Right, fPCB array patch coated with 100 µm Cu on the rear side. (i) Simulated temperature increases with or without device surface modification with 100 µm Cu. (a) Representative images of ChrimsonR expression in the targeted cortical regions from one mouse. The experiment was repeated independently on 8 animals. (b) Heatmap showing the spread of viral expression across cortical regions. (c) Example image showing the distribution of tdT, vglut1, and vgat transcripts in cortical column of somatosensory cortex limb representation. The experiment was repeated independently on 16 brain slices from 4 animals. (d) Schematic illustration of critical steps of surgical implantation. (a) Schematic of in vivo extracellular electrophysiology recordings in ChrimsonR-negative mice. (b) Left, raw recording traces showing the electrical artifacts during the stimulation period at different distances (1,000-5,000 μm) from μ-ILED. Right, processed traces using a custom-modified version of the Estimation and Removal of Array Artifacts via Sequential Principal Components Regression (ERAASR) algorithm. Lines and shaded areas represent mean ± s.e.m. (c) Schematic of the in vivo extracellular electrophysiology recording setup for evoked LFPs in ChrimsonR positive mice. (d) Traces showing evoked LFPs averaged across all electrodes of the MEA at different distances from the μ-ILED with varying input optical power. Lines and shaded areas represent mean ± s.e.m. (e) Summary of evoked LFP magnitudes at different distances from the μ-ILED under varying input optical power levels. Lines represent the mean from all electrodes, and dots indicate individual electrode measurements. (f) Same as (e), but for latency to valley in LFP waveforms. Dots represent individual electrodes; lines indicate the mean. (a) Number of sessions of 100 trials each across three levels of the task to reach the criterion of 80% success rate for animals expressing ChrimsonR (session median Level 1-12, Level 2-6, Level 3-12). Data are presented as mean ± s.e.m. Dots represent individual animals. (b) Summary data showing the total training length to complete Levels 1-3 and the number of sessions per day. Dark orange, group average; pale orange, individual trajectories; symbols mark the day individual animals reached criterion. (c) Timeline from water restriction to the end of training. Open-field locomotion was measured at the start of water restriction, before device implantation, and 2, 5, and 10 days post-implantation. (d) Left, average movement speed in an open-field arena across days, one-way ANOVA, F (1.739, 5.217) = 1.632, p = 0.2773. Middle, same as left, but for acceleration, one-way ANOVA, F (2.154, 6.461) = 2.358, p = 0.1693. Right, same as left, but for exploration rate, one-way ANOVA, F (1.497, 4.491) = 3.271, p = 0.1354. Box plots show median (line), 25th and 75th percentiles (bounds of box), minimum and maximum values (whiskers). (e) Scatter plots showing reaction times in all trials for the Level 3 task as a function of total cortical distance between stimulated digits. (f) Summary data of cumulative distributions of reaction times for all animals. (a) Line plots showing success rate as a function of spatial distance for 10 individual animals in the Level 3 task. The number of trials with correct or incorrect choices in each bin of spatial distance is plotted in the histogram. (b) Heatmap showing success rate for randomized non-target sequences grouped by specific stimulation locations at the first to the fourth stimulation digit in 10 ChrimsonR expressing animals. Red squares indicate the target stimulation sequence for each animal. (c) Two-sided Pearson's correlation analysis of spatial distance and success rate based on stimulation digit for individual animals. Dashed lines indicate 95% confidence intervals. (a) Left, schematic illustration of probing sequences with 75% (3 stimulation digits) and 50% similarity (2 stimulation digits) to the target sequences. Right, summary data showing the success rate of all probing sessions from all animals. 75% overlap: n = 75 sessions from 5 animals; 50% overlap: n = 41 sessions from 5 animals. (b) Left, schematic illustration of probing experiments with reversed sequences. Right, summary data showing success rate of all probing sessions from all animals, 0.7347 ± 0.00338; One sample t-test vs 0.5, p < 0.0001; n = 15 sessions from 5 animals. (c) Schematic illustrating single site stimulation task for mice to discriminate neighboring stimulation either on the same hemisphere or on the contralateral hemisphere. (d) Left, example of target (stimulation on a single cortical region) and non-target stimulation on a neighboring site. Middle, summary data showing discrimination performance on neighboring stimulation sites. Ipsilateral, 0.6151 ± 0.0168; one sample t-test vs 0.5, p < 0.0001; n = 58 sessions from 5 animals. Contralateral, 0.7081 ± 0.0257; one sample t-test vs 0.5, p < 0.0001; n = 20 sessions from 3 animals; Two-sided unpaired t-test, ipsi vs contra, p = 0.0053. Right, summary data showing discrimination of ipsilateral neighboring stimulation, 1st column (motor vs limb), 0.6788 ± 0.0542; one sample t-test vs 0.5, p = 0.0132. 2nd column (limb vs trunk), 0.6147 ± 0.0471; one sample t-test vs 0.5, p = 0.0288. 3rd column (trunk vs visual), 0.6007 ± 0.0149; one sample t-test vs 0.5, p < 0.0001. One-way ANOVA, Sidak's multiple comparisons test, 1st vs 2nd column, p = 0.5884; 1st vs 3rd column, p = 0.3287; 2nd vs 3rd column, p = 0.9792. (e) Left, example of a target and a non-target stimulation, where one digit was switched to a neighboring site. Middle, summary data showing no significant difference in discrimination performance when the switched neighboring digit was in the ipsilateral or contralateral hemisphere. Ipsilateral, 0.6132 ± 0.0239; one sample t-test vs 0.5, p < 0.0001; n = 24 sessions from 3 animals. Contralateral, 0.5935 ± 0.0249; one sample t-test vs 0.5, p = 0.0013; n = 20 sessions from 3 animals; Two-sided unpaired t-test, ipsi vs contra, p = 0.5742. Right, summary data showing discrimination performance when the switched digit was in the ipsilateral hemisphere, 1st column (motor vs limb), 0.5930 ± 0.0462; one sample t-test vs 0.5, p = 0.0786. 2nd column (limb vs trunk), 0.6600 ± 0.0620; one sample t-test vs 0.5, p = 0.0494. 3rd column (trunk vs visual), 0.6022 ± 0.0196; one sample t-test vs 0.5, p = 0.0008. One-way ANOVA, Sidak's multiple comparisons test, 1st vs 2nd column, p = 0.6523; 1st vs 3rd column, p = 0.9978; 2nd vs 3rd column, p = 0.7640. (f) Left, example of a target stimulation for mice to discriminate the target against each individual digit within it. Right, summary data of performance, 1st stim. (target vs 1st digit of the target), 0.6893 ± 0.0196; one sample t-test vs 0.5, p < 0.0001. 2nd stim., 0.7911 ± 0.0703; one sample t-test vs 0.5, p = 0.0033. 3rd stim., 0.7600 ± 0.0428; one sample t-test vs 0.5, p = 0.0017. 4th stim., 0.9020 ± 0.0132; one sample t-test vs 0.5, p < 0.0001. (g) Left, example of a target for mice to discriminate the target against its first digit, with the initiation site across all cortical regions. Right, summary data of performance, motor (target initiates from the motor cortex), 0.7000 ± 0.0397; one sample t-test vs 0.5, p = 0.0005. somalimb, 0.6355 ± 0.0539; one sample t-test vs 0.5, p = 0.0031. somatrunk, 0.6625 ± 0.0309; one sample t-test vs 0.5, p < 0.0001. visual, 0.7497 ± 0.0339; one sample t-test vs 0.5, p < 0.0001. One-way ANOVA, Sidak's multiple comparisons test, motor vs somalimb, p = 0.8672; motor vs somatrunk, p = 0.9840; motor vs visual, p = 0.9414; somalimb vs somatrunk, p = 0.9972; somalimb vs visual, p = 0.2442; somatrunk vs visual, p = 0.4324. (h) Left, example showing mice discriminate stimulations on the same site with different durations. Right, summary data of performance. 0.3 s (1.2 s target vs 0.3 s like-target), 0.7256 ± 0.0384; one sample t-test vs 0.5, p = 0.0004. 0.6 s, 0.7088 ± 0.0396; one sample t-test vs 0.5, p = 0.0012. 0.9 s, 0.6087 ± 0.0642; one sample t-test vs 0.5, p < 0.0001. All data are presented as mean ± s.e.m. Dots represent individual sessions of 100 trials. Supplementary Methods, Tables 1−4 and legends for Supplementary Videos 1−4. Dynamic spatial patterns generated by FOD with 8 × 8 µ-ILEDs. Independent control of µ-ILEDs and dynamic operation of FOD. Independent intensity control of FOD. Transient-state temperature distribution on the skull and brain surface. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. Wu, M., Yang, Y., Zhang, J. et al. Patterned wireless transcranial optogenetics generates artificial perception. Version of record: 08 December 2025 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing: Translational Research newsletter — top stories in biotechnology, drug discovery and pharma.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Electrocorticography uses non-penetrating electrodes embedded in flexible substrates to record electrical activity from the surface of the brain. To use the technology to develop minimally invasive, high-bandwidth brain–computer interfaces, it will be necessary to improve the number of recording channels and the scalability of devices, which could be achieved by merging electrodes and electronics onto a single substrate. Here we report a 50-μm-thick, mechanically flexible micro-electrocorticography brain–computer interface that integrates a 256 × 256 array of electrodes, signal processing, data telemetry and wireless powering on a single complementary metal–oxide–semiconductor substrate. The device contains 65,536 recording electrodes, from which we can simultaneously record a selectable subset of up to 1,024 channels at a given time. Our chip is wirelessly powered, and when implanted below the dura, it can communicate bidirectionally with an external relay station outside the body. We show that the device can provide chronic, reliable recordings for up to two weeks in pigs and up to two months in behaving non-human primates from the somatosensory, motor and visual cortices, decoding brain signals at high spatiotemporal resolution. This is a preview of subscription content, access via your institution Get Nature+, our best-value online-access subscription Subscribe to this journal Receive 12 digital issues and online access to articles Prices may be subject to local taxes which are calculated during checkout All electrophysiological data relevant to the figures presented in this paper are available via GitHub at https://github.com/klshepard/bisc with a version archived in Zenodo (https://doi.org/10.5281/zenodo.17074065)70. All other relevant data are available from the corresponding authors upon reasonable request. All scripts used for the data analysis are available via GitHub at https://github.com/klshepard/bisc. All other relevant codes are available from the corresponding authors upon reasonable request. A., Faggin, B. M., Votaw, S. & Oliveira, L. M. O. Reconstructing the engram: simultaneous, multisite, many single neuron recordings. Maynard, E. M., Nordhausen, C. T. & Normann, R. A. The Utah Intracortical Electrode Array: a recording structure for potential brain-computer interfaces. & Reichert, W. M. Response of brain tissue to chronically implanted neural electrodes. Salatino, J. W., Ludwig, K. A., Kozai, T. D. Y. & Purcell, E. K. Glial responses to implanted electrodes in the brain. Chronic recording capability of the Utah Intracortical Electrode Array in cat sensory cortex. Long-term stability of neural prosthetic control signals from silicon cortical arrays in rhesus macaque motor cortex. Volkova, K., Lebedev, M. A., Kaplan, A. & Ossadtchi, A. Decoding movement from electrocorticographic activity: a review. Consistency of long-term subdural electrocorticography in humans. Chronic subdural electrocorticography in nonhuman primates by an implantable wireless device for brain-machine interfaces. Chiang, C.-H. et al. Development of a neural interface for high-definition, long-term recording in rodents and nonhuman primates. Tchoe, Y. et al. Human brain mapping with multithousand-channel PtNRGrids resolves spatiotemporal dynamics. Kaiju, T. et al. High spatiotemporal resolution ECoG recording of somatosensory evoked potentials with flexible micro-electrode arrays. Comparison of decoding resolution of standard and high-density electrocorticogram electrodes. Duraivel, S. et al. High-resolution neural recordings improve the accuracy of speech decoding. Khodagholy, D. et al. NeuroGrid: recording action potentials from the surface of the brain. Steinmetz, N. A. et al. Neuropixels 2.0: a miniaturized high-density probe for stable, long-term brain recordings. Carmena, J. M. et al. Learning to control a brain–machine interface for reaching and grasping by primates. Hochberg, L. R. et al. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Brain micromotion around implants in the rodent somatosensory cortex. Biran, R., Martin, D. C. & Tresco, P. A. The brain tissue response to implanted silicon microelectrode arrays is increased when the device is tethered to the skull. Chronic, wireless recordings of large-scale brain activity in freely moving rhesus monkeys. Zhou, A. et al. A wireless and artefact-free 128-channel neuromodulation device for closed-loop stimulation and recording in non-human primates. Topalovic, U. et al. A wearable platform for closed-loop stimulation and recording of single-neuron and local field potential activity in freely moving humans. Oxley, T. J. et al. Minimally invasive endovascular stent-electrode array for high-fidelity, chronic recordings of cortical neural activity. Benabid, A. L. et al. An exoskeleton controlled by an epidural wireless brain–machine interface in a tetraplegic patient: a proof-of-concept demonstration. Musk, E. An integrated brain-machine interface platform with thousands of channels. Complications of deep brain stimulation surgery. Systematic review of hardware-related complications of deep brain stimulation: do new indications pose an increased risk? Zeng, N. et al. A wireless, mechanically flexible, 25 μm-thick, 65,536-channel subdural surface recording and stimulating microelectrode array with integrated antennas. IEEE Symposium on VLSI Technology and Circuits 1–2 (IEEE, 2023). Harrison, R. R. & Charles, C. A. low-power, low-noise CMOS amplifier for neural recording applications. Zhang, M. et al. Wireless compact neural interface for freely moving animal subjects: a review on wireless neural interface SoC designs. IEEE Solid-State Circuits Mag. & Humayun, M. S. In vitro electrical properties for iridium oxide versus titanium nitride stimulating electrodes. Raducanu, B. C. et al. Time multiplexed active neural probe with 1n356 parallel recording sites. IEEE Standard for Safety Levels with Respect to Human Exposure to Radio Frequency Electromagnetic Fields, 3 kHz to 300 GHz (IEEE, 2006). & Shepard, K. L. Bioelectronic devices: wirelessly powered implants. Kim, S., Tathireddy, P., Normann, R. A. & Solzbacher, F. Thermal impact of an active 3-D microelectrode array implanted in the brain. Physical principles for scalable neural recording. Craner, S. L. & Ray, R. H. Somatosensory cortex of the neonatal pig. I. Topographic organization of the primary somatosensory cortex (SI). & Xu, C. Comparison of MEG and EEG on the basis of somatic evoked responses elicited by stimulation of the snout in the juvenile swine. Sauleau, P., Lapouble, E., Val-Laillet, D. & Malbert, C. H. The pig model in brain imaging and neurosurgery. van der Maaten, L. aurens & Hinton, G. Visualizing data using t-SNE. Pollmann, E. H. et al. A subdural CMOS optical device for bidirectional neural interfacing. Analysis of dynamic brain imaging data. Stavisky, S. D., Kao, J. C., Nuyujukian, P., Ryu, S. I. & Shenoy, K. V. A high performing brain–machine interface driven by low-frequency local field potentials alone and together with spikes. Spectrum estimation and harmonic analysis. Wold, S., Sjöström, M. & Eriksson, L. PLS-regression: a basic tool of chemometrics. Das, A., Zabeh, E., Ermentrout, B. & Jacobs, J. Planar, spiral, and concentric traveling waves distinguish cognitive states in human memory. Preprint at bioRxiv https://doi.org/10.1101/2024.01.26.577456 (2024). Zabeh, E., Foley, N. C., Jacobs, J. & Gottlieb, J. P. Beta traveling waves in monkey frontal and parietal areas encode recent reward history. Deng, J. et al. ImageNet: a large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition 248–255 (IEEE, 2009). Deep learning-driven characterization of single cell tuning in primate visual area V4 unveils topological organization. Preprint at bioRxiv https://doi.org/10.1101/2023.05.12.540591 (2023). Jia, X., Tanabe, S. & Kohn, A. Gamma and the coordination of spiking activity in early visual cortex. Woo, S. et al. ConvNeXt V2: co-designing and scaling ConvNets with masked autoencoders. IEEE/CVF Conference on Computer Vision and Pattern Recognition 16133–16142 (IEEE, 2023). Pierzchlewicz, P. et al. Energy guided diffusion for generating neurally exciting images. Diverse task-driven modeling of macaque V4 reveals functional specialization towards semantic tasks. Inception loops discover what excites neurons most using deep predictive models. Fu, J. et al. Pattern completion and disruption characterize contextual modulation in mouse visual cortex. Preprint at bioRxiv https://doi.org/10.1101/2023.03.13.532473 (2023). Franke, K. et al. State-dependent pupil dilation rapidly shifts visual feature selectivity. Cowley, B. R., Stan, P. L., Pillow, J. W. & Smith, M. A. Compact deep neural network models of visual cortex. Preprint at bioRxiv https://doi.org/10.1101/2023.11.22.568315 (2023). Bashivan, P., Kar, K. & DiCarlo, J. J. Neural population control via deep image synthesis. Zhang, H., Watrous, A. J., Patel, A. & Jacobs, J. Theta and alpha oscillations are traveling waves in the human neocortex. Das, A. et al. Spontaneous neuronal oscillations in the human insula are hierarchically organized traveling waves. Highly ordered arrangement of single neurons in orientation pinwheels. Kara, P. & Boyd, J. D. A micro-architecture for binocular disparity and ocular dominance in visual cortex. Willett, F. R. et al. A high-performance speech neuroprosthesis. Metzger, S. L. et al. A high-performance neuroprosthesis for speech decoding and avatar control. & Sokal, A. D. Class E—a new class of high-efficiency tuned single-ended switching power amplifiers. An adaptive impedance-matching network based on a novel capacitor matrix for wireless power transfer. Electronic and ionic materials for neurointerfaces. Jung, T. et al. klshepard/bisc: bioelectronic interface system to the cortex (a wireless subdural-contained 65,536-electrode, 1,024-channel brain-computer interface). Calabrese, E. et al. A diffusion tensor MRI atlas of the postmortem rhesus macaque brain. Bakker, R., Tiesinga, P. & Kötter, R. The Scalable Brain Atlas: instant web-based access to public brain atlases and related content. This work was partly supported by the Defense Advanced Research Project Agency (DARPA) under contract number N66001-17-C-4001, the Department of the Defense Congressionally Directed Medical Research Program under contract number HT9425-23-1-0758, the National Science Foundation under grant number 1546296 and the National Institutes of Health under grant number R01DC019498. We acknowledge the use of facilities and instrumentation at the Columbia Nano Initiative, the CUNY ASRC and the UPenn Quattrone Nanofabrication Facility. We also thank Y. Borisenkov, A. Banees and K. Kim at Columbia University for help with chip processing and many helpful discussions. These authors contributed equally: Taesung Jung, Nanyu Zeng. Department of Electrical Engineering, Columbia University, New York, NY, USA Taesung Jung, Nanyu Zeng, Jason D. Fabbri, Rizwan Huq, Mohit Sharma, Yaoxing Hu, Girish Ramakrishnan, Kevin Tien, Abhinav Parihar, Heyu Yin, Ilke Uguz & Kenneth L. Shepard Kampto Neurotech LLC, Troy, NY, USA Department of Computer Science, Columbia University, New York, NY, USA Guy Eichler, Paolo Mantovani, Alexander Misdorp & Luca P. Carloni Department of Ophthalmology, Byers Eye Institute, Stanford University, Stanford, CA, USA Zhe Li, Konstantin Willeke, Gabrielle J. Rodriguez, Cate Nealley, Saumil Patel & Andreas Tolias Zhe Li, Konstantin Willeke, Gabrielle J. Rodriguez, Cate Nealley, Sophia Sanborn, Saumil Patel & Andreas Tolias Wu Tsai Neurosciences Institute, Stanford University, Stanford, CA, USA Zhe Li, Konstantin Willeke, Gabrielle J. Rodriguez, Cate Nealley, Sophia Sanborn, Saumil Patel & Andreas Tolias Department of Biomedical Engineering, Columbia University, New York, NY, USA Erfan Zabeh, Anup Das & Kenneth L. Shepard Institute of Computer Science and Campus Institute Data Science, University of Göttingen, Göttingen, Germany Center for Neural Science, New York University, New York, NY, USA Katie E. Wingel, Agrita Dubey & Bijan Pesaran Department of Neurosurgery, University of Pennsylvania, Philadelphia, PA, USA Katie E. Wingel, Agrita Dubey, Denise Oswalt, Daniel Yoshor & Bijan Pesaran Department of Neuroscience, University of Pennsylvania, Philadelphia, PA, USA Denise Oswalt & Bijan Pesaran Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, USA Denise Oswalt & Bijan Pesaran Center for Neuroscience and Artificial Intelligence, Department of Neuroscience, Baylor College of Medicine, Houston, TX, USA Tori Shinn & Andreas Tolias Neuroscience Research Institute, University of California Santa Barbara, Santa Barbara, CA, USA Tjitse van der Molen & Kenneth S. Kosik Department of Molecular, Cellular and Developmental Biology, University of California Santa Barbara, Santa Barbara, CA, USA Tjitse van der Molen & Kenneth S. Kosik Department of Neurological Surgery, Columbia University, New York, NY, USA Ian Gonzales, Eleonora Spinazzi, Brett Youngerman & Kenneth L. Shepard Department of Applied Physics, Caltech, Pasadena, CA, USA Department of Physics, Caltech, Pasadena, CA, USA Department of Bioengineering, Caltech, Pasadena, CA, USA Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, USA Department of Pathology and Cell Biology, Columbia University, New York, NY, USA Department of Neurology and Neuroscience Institute, University of Chicago, Chicago, IL, USA Shirley Ryan Ability Labs, Chicago, IL, USA Department of Physical Medicine and Rehabilitation, Northwestern University, Chicago, IL, USA Department of Electrical Engineering, Stanford University, Stanford, CA, USA Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar designed the implant circuitry. and T.J. implemented the relay station hardware. implemented the relay station software. performed the bench-top characterizations. performed the in vivo experiments on the porcine subject. conducted the porcine data analysis and histology. and T.J. performed the in vivo experiments on the motor cortex of the NHP. performed the motor cortex data analysis. performed the in vivo experiments on the visual cortex of the NHP. Z.L., K.W., A.T., S.P., D.O., R.J.C., E.Z., A. Das and J.J. performed the visual cortex data analysis. wrote the paper with review and editing contributed by all authors. Correspondence to Andreas Tolias or Kenneth L. Shepard. is a principal with Kampto Neurotech, LLC, which is commercializing the BISC technology. The BISC technology is patented under US patent 11617890, issued on 4 April 2023, and exclusively licensed to Kampto from Columbia University. The other authors declare no competing interests. Nature Electronics thanks the anonymous reviewers for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. (a) Electrochemical impedance spectroscopy of titanium nitride electrode. (b) Frequency response across different gain configurations from a representative 16×16 recording. Note that gain is programmed through a single back-end amplifier that is shared by all pixels. Error bars indicate standard error (SE), and dashed rectangle marks the effects of boxcar sampling (flat band gains: 53.7 ± 0.20 dB, 57.2 ± 0.21 dB, 60.7 ± 0.20 dB, 64.2 ± 0.19 dB, values: mean ± SD. (c) Histogram of normalized channel gain variation for each recording mode (16×16 mode: 100 ± 5.1%, 32×32 mode: 100 ± 4.8%, values: mean ± SD. (d) Frequency response across different high-pass (HP) filter configurations from a representative 16×16 recording. (e) Input-referred noise (IRN) spectrum averaged over representative pixels (n = 10) for each recording mode. (f) Histogram of channel IRN for each recording mode, integrated from 10 Hz to 4 kHz (16×16 mode: 7.68 ± 3.11 μVRMS, 32×32 mode: 16.51 ± 6.85 μVRMS, values: mean ± SD. (a) We presented static colored natural images, while the monkey maintained fixation (120 ms presentation time per image, 15 images per trial, 1200 ms inter-trial period). Each image (10°×10°) was centered 3° to the right and below the fixation spot. (b) Model architecture: Pre-processed stimuli (184 × 184 pixels) and neuronal responses were used to train a neural predictive model, which takes images as an input and outputs an estimate of the underlying neuronal activity. We passed the images through a ConvNext model, pre-trained on an image classification task to obtain image embeddings, that is a shared feature space. We then computed the neuronal responses by passing the feature activations through a transformer-based readout followed by a non-linearity stage. (c) Explainable variance, a measure of response reliability to natural images, plotted against the model's predictive performance (correlation between prediction and average neural response to repeated presentations) of all 144 channels (explainable variance 0.24 ± 0.09, and correlation to average 0.69 ± 0.14. values: mean ± SD). Only channels with an explainable variance greater than or equal to 0.1 are included in these analyses. (d) Spatial map of explainable variance across the recording array (same layout as in Fig. (e) Same as (d), but showing the model's predictive performance (correlation to average neural response). (f) Schematic illustrating optimization of maximally exciting images (MEIs). A random starting image was iteratively optimized to elicit maximal activity for each in-silico channel, revealing the visual features to which that channel is selective. Three example MEIs from areas V1, V2, and V4 are shown. (g) MEIs for all 144 channels across the array which reliably responded to repeated image presentations. MEIs in area V1 are characterized by oriented Gabor filters, while the channels overlying area V2 and V4 exhibit more complex, color opponent feature tuning. Credit: cow image in a,b, Nicolas Vigier, flickr under a Creative Commons license CC0. Normalized somatosensory evoked potential (SSEP) recording from a porcine model, trial averaged (n = 100 per location). Motor cortex recording from a NHP model performing asynchronous reach-and-grab task. Dot-triggered-average responses of all channels without filtering. Dot-triggered-average responses of all channels after wavelet transformation (central frequency 8 Hz). Dot-triggered-average responses of all channels after wavelet transformation (central frequency 16 Hz). Dot-triggered-average responses of all channels after wavelet transformation (central frequency 32 Hz). Dot-triggered-average responses of all channels after wavelet transformation (central frequency 64 Hz). Dot-triggered-average responses of all channels after wavelet transformation (central frequency 128 Hz). Dot-triggered travelling waves used for decoding stimuli location. The travelling waves are computed from the γ-band (30–90 Hz) signals recorded from 32 × 32 spatially dense channels at a pitch of 26.5 μm × 29 μm. The spatiotemporal sequence of these travelling waves, measured within each dot presentation, is used to predict the current location of the dot stimuli presented to the subject. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. et al. A wireless subdural-contained brain–computer interface with 65,536 electrodes and 1,024 channels. Version of record: 08 December 2025 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.