Pub Date : 2025-09-12DOI: 10.1038/s43588-025-00866-x
Tao Yan, Yanchen Guo, Tiankuang Zhou, Guocheng Shao, Shanglong Li, Ruqi Huang, Qionghai Dai, Lu Fang
The field of photonic neural networks has experienced substantial growth, driven by its potential to enable ultrafast artificial intelligence inference and address the escalating demand for computing speed and energy efficiency. However, realizing nonlinearity-complete all-optical neurons is still challenging, constraining the performance of photonic neural networks. Here we report a complete photonic integrated neuron (PIN) with spatiotemporal feature learning capabilities and reconfigurable structures for nonlinear all-optical computing. By interleaving the spatiotemporal dimension of photons and leveraging the Kerr effect, PIN performs high-order temporal convolution and all-optical nonlinear activation monolithically on a silicon-nitride photonic chip, achieving neuron completeness of weighted interconnects and nonlinearities. We develop the PIN chip system and demonstrate its remarkable performance in high-accuracy image classification and human motion generation. PIN enables ultrafast spatialtemporal processing with a latency as low as 240 ps, paving the way for advancing machine intelligence into the subnanosecond regime. This study reports a complete photonic neuron integrated on a silicon-nitride chip, enabling ultrafast all-optical computing with nonlinear multi-kernel convolution for image recognition and motion generation.
{"title":"A complete photonic integrated neuron for nonlinear all-optical computing","authors":"Tao Yan, Yanchen Guo, Tiankuang Zhou, Guocheng Shao, Shanglong Li, Ruqi Huang, Qionghai Dai, Lu Fang","doi":"10.1038/s43588-025-00866-x","DOIUrl":"10.1038/s43588-025-00866-x","url":null,"abstract":"The field of photonic neural networks has experienced substantial growth, driven by its potential to enable ultrafast artificial intelligence inference and address the escalating demand for computing speed and energy efficiency. However, realizing nonlinearity-complete all-optical neurons is still challenging, constraining the performance of photonic neural networks. Here we report a complete photonic integrated neuron (PIN) with spatiotemporal feature learning capabilities and reconfigurable structures for nonlinear all-optical computing. By interleaving the spatiotemporal dimension of photons and leveraging the Kerr effect, PIN performs high-order temporal convolution and all-optical nonlinear activation monolithically on a silicon-nitride photonic chip, achieving neuron completeness of weighted interconnects and nonlinearities. We develop the PIN chip system and demonstrate its remarkable performance in high-accuracy image classification and human motion generation. PIN enables ultrafast spatialtemporal processing with a latency as low as 240 ps, paving the way for advancing machine intelligence into the subnanosecond regime. This study reports a complete photonic neuron integrated on a silicon-nitride chip, enabling ultrafast all-optical computing with nonlinear multi-kernel convolution for image recognition and motion generation.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 12","pages":"1202-1213"},"PeriodicalIF":18.3,"publicationDate":"2025-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145056517","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-12DOI: 10.1038/s43588-025-00856-z
Jonah Rosenblum, Juechu Dong, Satish Narayanasamy
Genomic data from a single institution lacks global diversity representation, especially for rare variants and diseases. Confidential computing can enable collaborative genome-wide association studies (GWAS) without compromising privacy or accuracy. However, due to limited secure memory space and performance overheads, previous solutions fail to support widely used regression methods. Here we present SECRET-GWAS—a rapid, privacy-preserving, population-scale, collaborative GWAS tool. We discuss several system optimizations, including streaming, batching, data parallelization and reducing trusted hardware overheads to efficiently scale linear and logistic regression to over a thousand processor cores on an Intel SGX-based cloud platform. In addition, we protect SECRET-GWAS against several hardware side-channel attacks. SECRET-GWAS is an open-source tool and works with the widely used Hail genomic analysis framework. Our experiments on Azure’s Confidential Computing platform demonstrate that SECRET-GWAS enables multivariate linear and logistic regression GWAS queries on population-scale datasets from ten independent sources in just 4.5 and 29 minutes, respectively. Secure collaborative genome-wide association studies (GWAS) with population-scale datasets address gaps in genomic data. This work proposes SECRET-GWAS and system optimizations that overcome resource constraints and exploit parallelism, while maintaining privacy and accuracy.
{"title":"Confidential computing for population-scale genome-wide association studies with SECRET-GWAS","authors":"Jonah Rosenblum, Juechu Dong, Satish Narayanasamy","doi":"10.1038/s43588-025-00856-z","DOIUrl":"10.1038/s43588-025-00856-z","url":null,"abstract":"Genomic data from a single institution lacks global diversity representation, especially for rare variants and diseases. Confidential computing can enable collaborative genome-wide association studies (GWAS) without compromising privacy or accuracy. However, due to limited secure memory space and performance overheads, previous solutions fail to support widely used regression methods. Here we present SECRET-GWAS—a rapid, privacy-preserving, population-scale, collaborative GWAS tool. We discuss several system optimizations, including streaming, batching, data parallelization and reducing trusted hardware overheads to efficiently scale linear and logistic regression to over a thousand processor cores on an Intel SGX-based cloud platform. In addition, we protect SECRET-GWAS against several hardware side-channel attacks. SECRET-GWAS is an open-source tool and works with the widely used Hail genomic analysis framework. Our experiments on Azure’s Confidential Computing platform demonstrate that SECRET-GWAS enables multivariate linear and logistic regression GWAS queries on population-scale datasets from ten independent sources in just 4.5 and 29 minutes, respectively. Secure collaborative genome-wide association studies (GWAS) with population-scale datasets address gaps in genomic data. This work proposes SECRET-GWAS and system optimizations that overcome resource constraints and exploit parallelism, while maintaining privacy and accuracy.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"825-835"},"PeriodicalIF":18.3,"publicationDate":"2025-09-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145056513","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-10DOI: 10.1038/s43588-025-00871-0
A benchmark — MaCBench — is developed for evaluating the scientific knowledge of vision language models (VLMs). Evaluation of leading VLMs reveals that they excel at basic scientific tasks such as equipment identification, but struggle with spatial reasoning and multistep analysis — a limitation for autonomous scientific discovery.
{"title":"Vision language models excel at perception but struggles with scientific reasoning","authors":"","doi":"10.1038/s43588-025-00871-0","DOIUrl":"10.1038/s43588-025-00871-0","url":null,"abstract":"A benchmark — MaCBench — is developed for evaluating the scientific knowledge of vision language models (VLMs). Evaluation of leading VLMs reveals that they excel at basic scientific tasks such as equipment identification, but struggle with spatial reasoning and multistep analysis — a limitation for autonomous scientific discovery.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 10","pages":"852-853"},"PeriodicalIF":18.3,"publicationDate":"2025-09-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145034850","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-09DOI: 10.1038/s43588-025-00839-0
Jun Yin (, ), Honghao Chen (, ), Jiangjie Qiu (, ), Wentao Li (, ), Peng He (, ), Jiali Li (, ), Iftekhar A. Karimi, Xiaocheng Lan (, ), Tiefeng Wang (, ), Xiaonan Wang (, )
With approximately 90% of industrial reactions occurring on surfaces, the role of heterogeneous catalysts is paramount. Currently, accurate surface exposure prediction is vital for heterogeneous catalyst design, but it is hindered by the high costs of experimental and computational methods. Here we introduce a foundation force-field-based model for predicting surface exposure and synthesizability (SurFF) across intermetallic crystals, which are essential materials for heterogeneous catalysts. We created a comprehensive intermetallic surface database using an active learning method and high-throughput density functional theory calculations, encompassing 12,553 unique surfaces and 344,200 single points. SurFF achieves density-functional-theory-level precision with a prediction error of 3 meV Å−2 and enables large-scale surface exposure prediction with a 105-fold acceleration. Validation against computational and experimental data both show strong alignment. We applied SurFF for large-scale predictions of surface energy and Wulff shapes for over 6,000 intermetallic crystals, providing valuable data for the community. A foundation machine learning model, SurFF, enables DFT-accurate predictions of surface energies and morphologies in intermetallic catalysts, achieving over 105-fold acceleration for high-throughput materials screening.
{"title":"SurFF: a foundation model for surface exposure and morphology across intermetallic crystals","authors":"Jun Yin \u0000 (, ), Honghao Chen \u0000 (, ), Jiangjie Qiu \u0000 (, ), Wentao Li \u0000 (, ), Peng He \u0000 (, ), Jiali Li \u0000 (, ), Iftekhar A. Karimi, Xiaocheng Lan \u0000 (, ), Tiefeng Wang \u0000 (, ), Xiaonan Wang \u0000 (, )","doi":"10.1038/s43588-025-00839-0","DOIUrl":"10.1038/s43588-025-00839-0","url":null,"abstract":"With approximately 90% of industrial reactions occurring on surfaces, the role of heterogeneous catalysts is paramount. Currently, accurate surface exposure prediction is vital for heterogeneous catalyst design, but it is hindered by the high costs of experimental and computational methods. Here we introduce a foundation force-field-based model for predicting surface exposure and synthesizability (SurFF) across intermetallic crystals, which are essential materials for heterogeneous catalysts. We created a comprehensive intermetallic surface database using an active learning method and high-throughput density functional theory calculations, encompassing 12,553 unique surfaces and 344,200 single points. SurFF achieves density-functional-theory-level precision with a prediction error of 3 meV Å−2 and enables large-scale surface exposure prediction with a 105-fold acceleration. Validation against computational and experimental data both show strong alignment. We applied SurFF for large-scale predictions of surface energy and Wulff shapes for over 6,000 intermetallic crystals, providing valuable data for the community. A foundation machine learning model, SurFF, enables DFT-accurate predictions of surface energies and morphologies in intermetallic catalysts, achieving over 105-fold acceleration for high-throughput materials screening.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"782-792"},"PeriodicalIF":18.3,"publicationDate":"2025-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145031334","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-09DOI: 10.1038/s43588-025-00848-z
Ilia Sucholutsky, Katherine M. Collins, Nori Jacoby, Bill D. Thompson, Robert D. Hawkins
Large language models (LLMs) are already transforming the study of individual cognition, but their application to studying collective cognition has been underexplored. We lay out how LLMs may be able to address the complexity that has hindered the study of collectives and raise possible risks that warrant new methods.
{"title":"Using LLMs to advance the cognitive science of collectives","authors":"Ilia Sucholutsky, Katherine M. Collins, Nori Jacoby, Bill D. Thompson, Robert D. Hawkins","doi":"10.1038/s43588-025-00848-z","DOIUrl":"10.1038/s43588-025-00848-z","url":null,"abstract":"Large language models (LLMs) are already transforming the study of individual cognition, but their application to studying collective cognition has been underexplored. We lay out how LLMs may be able to address the complexity that has hindered the study of collectives and raise possible risks that warrant new methods.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"704-707"},"PeriodicalIF":18.3,"publicationDate":"2025-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145031329","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-08DOI: 10.1038/s43588-025-00846-1
Yu Zheng, Fengli Xu, Yuming Lin, Paolo Santi, Carlo Ratti, Qi R. Wang, Yong Li
City plans are the product of integrating human creativity with emerging technologies, which continuously evolve and reshape urban morphology and environments. Here we argue that large language models hold large untapped potential in addressing the growing complexities of urban planning and enabling a more holistic, innovative and responsive approach to city design. By harnessing their advanced generation and simulation capabilities, large language models can contribute as an intelligent assistant for human planners in synthesizing conceptual ideas, generating urban designs and evaluating the outcomes of planning efforts. Large language models remain largely unexplored is the design of cities. In this Perspective, the authors discuss the potential opportunities brought by these models in assisting urban planning.
{"title":"Urban planning in the era of large language models","authors":"Yu Zheng, Fengli Xu, Yuming Lin, Paolo Santi, Carlo Ratti, Qi R. Wang, Yong Li","doi":"10.1038/s43588-025-00846-1","DOIUrl":"10.1038/s43588-025-00846-1","url":null,"abstract":"City plans are the product of integrating human creativity with emerging technologies, which continuously evolve and reshape urban morphology and environments. Here we argue that large language models hold large untapped potential in addressing the growing complexities of urban planning and enabling a more holistic, innovative and responsive approach to city design. By harnessing their advanced generation and simulation capabilities, large language models can contribute as an intelligent assistant for human planners in synthesizing conceptual ideas, generating urban designs and evaluating the outcomes of planning efforts. Large language models remain largely unexplored is the design of cities. In this Perspective, the authors discuss the potential opportunities brought by these models in assisting urban planning.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"727-736"},"PeriodicalIF":18.3,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145024878","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-08DOI: 10.1038/s43588-025-00865-y
Pedro Burgos
While leading tech companies race to build ever-larger models, researchers in Brazil, India and Africa are using clever tricks to remix big labs’ LLMs to bring AI to billions of users.
{"title":"The other AI revolution: how the Global South is building and repurposing language models that speak to billions","authors":"Pedro Burgos","doi":"10.1038/s43588-025-00865-y","DOIUrl":"10.1038/s43588-025-00865-y","url":null,"abstract":"While leading tech companies race to build ever-larger models, researchers in Brazil, India and Africa are using clever tricks to remix big labs’ LLMs to bring AI to billions of users.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"691-694"},"PeriodicalIF":18.3,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145024894","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-08DOI: 10.1038/s43588-025-00860-3
Yudeng Lin, Jianshi Tang
A recent study demonstrates the potential of using in-memory computing architecture for implementing large language models for an improved computational efficiency in both time and energy while maintaining a high accuracy.
{"title":"Overcoming computational bottlenecks in large language models through analog in-memory computing","authors":"Yudeng Lin, Jianshi Tang","doi":"10.1038/s43588-025-00860-3","DOIUrl":"10.1038/s43588-025-00860-3","url":null,"abstract":"A recent study demonstrates the potential of using in-memory computing architecture for implementing large language models for an improved computational efficiency in both time and energy while maintaining a high accuracy.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"711-712"},"PeriodicalIF":18.3,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145024835","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-08DOI: 10.1038/s43588-025-00854-1
Nathan Leroux, Paul-Philipp Manea, Chirag Sudarshan, Jan Finkbeiner, Sebastian Siegel, John Paul Strachan, Emre Neftci
Transformer networks, driven by self-attention, are central to large language models. In generative transformers, self-attention uses cache memory to store token projections, avoiding recomputation at each time step. However, graphics processing unit (GPU)-stored projections must be loaded into static random-access memory for each new generation step, causing latency and energy bottlenecks. Here we present a custom self-attention in-memory computing architecture based on emerging charge-based memories called gain cells, which can be efficiently written to store new tokens during sequence generation and enable parallel analog dot-product computation required for self-attention. However, the analog gain-cell circuits introduce non-idealities and constraints preventing the direct mapping of pre-trained models. To circumvent this problem, we design an initialization algorithm achieving text-processing performance comparable to GPT-2 without training from scratch. Our architecture reduces attention latency and energy consumption by up to two and four orders of magnitude, respectively, compared with GPUs, marking a substantial step toward ultrafast, low-power generative transformers. Leveraging in-memory computing with emerging gain-cell devices, the authors accelerate attention—a core mechanism in large language models. They train a 1.5-billion-parameter model, achieving up to a 70,000-fold reduction in energy consumption and a 100-fold speed-up compared with GPUs.
{"title":"Analog in-memory computing attention mechanism for fast and energy-efficient large language models","authors":"Nathan Leroux, Paul-Philipp Manea, Chirag Sudarshan, Jan Finkbeiner, Sebastian Siegel, John Paul Strachan, Emre Neftci","doi":"10.1038/s43588-025-00854-1","DOIUrl":"10.1038/s43588-025-00854-1","url":null,"abstract":"Transformer networks, driven by self-attention, are central to large language models. In generative transformers, self-attention uses cache memory to store token projections, avoiding recomputation at each time step. However, graphics processing unit (GPU)-stored projections must be loaded into static random-access memory for each new generation step, causing latency and energy bottlenecks. Here we present a custom self-attention in-memory computing architecture based on emerging charge-based memories called gain cells, which can be efficiently written to store new tokens during sequence generation and enable parallel analog dot-product computation required for self-attention. However, the analog gain-cell circuits introduce non-idealities and constraints preventing the direct mapping of pre-trained models. To circumvent this problem, we design an initialization algorithm achieving text-processing performance comparable to GPT-2 without training from scratch. Our architecture reduces attention latency and energy consumption by up to two and four orders of magnitude, respectively, compared with GPUs, marking a substantial step toward ultrafast, low-power generative transformers. Leveraging in-memory computing with emerging gain-cell devices, the authors accelerate attention—a core mechanism in large language models. They train a 1.5-billion-parameter model, achieving up to a 70,000-fold reduction in energy consumption and a 100-fold speed-up compared with GPUs.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"813-824"},"PeriodicalIF":18.3,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.nature.comhttps://www.nature.com/articles/s43588-025-00854-1.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145024889","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-01DOI: 10.1038/s43588-025-00859-w
An integrated platform, Digital Twin for Chemical Science (DTCS), is developed to connect first-principles theory with spectroscopic measurements through a bidirectional feedback loop. By predicting and refining chemical reaction mechanisms before, during and after experiments, DTCS enables the interpretation of spectra and supports real-time decision-making in chemical characterization.
{"title":"A digital twin that interprets and refines chemical mechanisms","authors":"","doi":"10.1038/s43588-025-00859-w","DOIUrl":"10.1038/s43588-025-00859-w","url":null,"abstract":"An integrated platform, Digital Twin for Chemical Science (DTCS), is developed to connect first-principles theory with spectroscopic measurements through a bidirectional feedback loop. By predicting and refining chemical reaction mechanisms before, during and after experiments, DTCS enables the interpretation of spectra and supports real-time decision-making in chemical characterization.","PeriodicalId":74246,"journal":{"name":"Nature computational science","volume":"5 9","pages":"713-714"},"PeriodicalIF":18.3,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144981688","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：