AlphaFold for Materials Science Has Arrived! Microsoft’s AI Model Featured in Nature

In fact, Microsoft launched the MatterGen model a year ago—around the same time that Google DeepMind’s GNoME tool successfully predicted 2.2 million crystal structures, sparking widespread academic discussion. Compared to its predecessor, co-author Tian Xie of this new study claims that MatterGen has undergone a major evolution. Overall, its performance metrics have been upgraded: not only has it successfully synthesized novel materials, but also the deviation between predicted and target property values has been reduced to within 20%. Moreover, thanks to a newly developed structural matching algorithm, the proportion of stable and novel materials generated by MatterGen has more than doubled. In addition, Microsoft cautioned in its official blog: MatterGen works best when paired with another of its models—MatterSim. In the field of materials discovery, these two models play distinct yet complementary roles, reshaping how researchers design and validate new materials. In short: MatterGen proposes candidate structures; MatterSim further predicts their feasibility. Once MatterGen generates potential material structures, MatterSim applies rigorous computational analysis to predict their viability—acting like a sieve, distinguishing theoretically possible candidates from those practically realizable. Reportedly, the MatterSim model integrates deep learning techniques to simulate atomic interactions. It can simulate diverse materials—including metals, oxides, sulfides, halides—and various states (e.g., crystalline, amorphous solids, and liquids) across temperature ranges from absolute zero to 5,000 Kelvin and pressure ranges from standard atmospheric pressure up to ten million times atmospheric pressure. A customized MatterSim model achieves experimental-level simulation accuracy using only 3% of the original training data. So what are MatterGen’s practical applications? It synthesized a novel material: TaCr₂O₆. The paper highlights several representative use cases. For instance, generating materials within targeted chemical systems. Compared against substitution methods and random structure search (RSS), MatterGen achieves the highest proportion of S.U.N. (Stable, Unique, Novel) structures across multiple chemical systems. Its performance is especially outstanding in complex five-element systems, demonstrating clear computational efficiency advantages—and discovering more novel structures even in the V–Sr–O (vanadium–strontium–oxygen) system. Another example is designing materials with targeted magnetic, electronic, and mechanical properties. As shown, for inverse-design tasks related to magnetic, electronic, and mechanical properties, MatterGen consistently steers generated material property values toward user-specified targets—even under varying amounts of labeled training data. Moreover, under limited DFT computation budgets, MatterGen identifies significantly more S.U.N. materials satisfying extreme performance constraints than conventional screening approaches. Finally, the paper also discusses designing magnets with low supply-chain risk. Briefly put, taking high-magnetic-density magnets with low supply-chain risk as an example: MatterGen generates structures distributed around the target values; when jointly optimizing for magnetic density and low Herfindahl–Hirschman Index (HHI)—a metric reflecting supply-chain concentration—it reduces reliance on high-risk elements while rediscovering many experimentally synthesized permanent-magnet-like structures.

Who Said Transformers Can’t Handle Multi-Scale Features? CAS, Peking University, Microsoft Research Asia, and Baidu Jointly Propose HRFormer—Reducing Memory and Parameters by 40% | NeurIPS 2021

[New Intelligence Digest] Transformers have shone brightly in computer vision (CV); how to integrate them with traditional CNNs—and leverage CNN strengths—has long been a key research question. Recently, researchers from the Chinese Academy of Sciences (CAS), Peking University, Microsoft Research Asia, and Baidu jointly proposed the HRFormer model, capable of processing images at multiple resolutions in parallel. This new architecture reduces parameter count by 40% while delivering superior performance! Vision Transformer (ViT) demonstrates exceptional performance on ImageNet classification tasks. Building upon ViT, numerous follow-up works have improved classification accuracy via knowledge distillation, deeper network architectures, or incorporating convolutional operations. Other research directions focus on extending Transformers to broader vision tasks—such as object detection, semantic segmentation, and video understanding. ViT divides input images into sequences of fixed-size patches (e.g., 16×16 pixels) and extracts feature representations for each patch. A key limitation of this approach is that ViT’s representation space loses substantial fine-grained detail—leading to suboptimal performance on tasks requiring precise spatial prediction. Moreover, ViT produces only single-scale feature representations, lacking the ability to capture multi-scale transformations commonly achieved through data augmentation. To mitigate feature granularity loss and better model multi-scale variations, researchers proposed the High-Resolution Transformer (HRFormer)—a Transformer variant enriched with richer spatial information—and constructed a multi-resolution representation framework tailored for dense prediction. To reduce feature granularity loss and model multi-scale variations, researchers from the Institute of Computing Technology (ICT) at CAS, Peking University, Microsoft Research Asia, and Baidu jointly proposed HRFormer—a high-resolution Transformer embedding richer spatial information—capable of constructing multi-resolution representations for dense prediction. The paper has been accepted by NeurIPS 2021. Paper co-author Jingdong Wang serves as Chief CV Architect at Baidu AI. He was a Senior Principal Researcher at Microsoft Research Asia from September 2007 to September 2021. His primary research areas include neural architecture design, human pose estimation, semantic segmentation, image classification, object detection, large-scale indexing, and salient object detection. Another co-author, Dr. Xilin Chen, is a researcher at ICT-CAS, and a Fellow of ACM, IEEE, IAPR, and the China Computer Federation. He is also a recipient of the National Science Fund for Distinguished Young Scholars. His main research interests lie in computer vision, pattern recognition, multimedia technologies, and multimodal human–computer interfaces. He has led numerous major and key projects funded by the National Natural Science Foundation of China, the National Basic Research Program of China (973 Program), and others. HRFormer adopts the multi-resolution parallel design from HRNet. Specifically, HRFormer employs convolutional layers in both the stem and the first stage—since prior research has empirically confirmed that convolutions perform better during early-stage image processing.

CAS Collaborates with Microsoft to Enable Faster Scientific Innovation

3 days ago

On October 20, CAS announced its integration framework is now available in Microsoft Copilot, granting users access to trusted scientific content and the specialized CAS Newton AI tool. This partnership aims to streamline scientific workflows by providing reliable data within popular AI platforms.

cen

2026-10-29

CAS Collaborates with Microsoft to Enable Faster Scientific Innovation

CAS Collaborates with Microsoft to Enable Faster Scientific Innovation

FAQ