Shanghai AI Lab Launches MatVerse Multimodal Materials Science Model

At a “New Tiangong Kaiwu” science and technology achievement event held by the National Science and Technology Dissemination Center, Suzhou National Laboratory, the Shanghai Artificial Intelligence Laboratory and Shanghai Jiao Tong University introduced MatVerse, a multimodal large model and agent application system for materials science. The announcement, reported on October 28, 2026, by the Shanghai Artificial Intelligence Laboratory, positioned MatVerse as a step from scientific question answering toward hypothesis-driven materials discovery. The model is designed to process heterogeneous scientific data, including X-ray diffraction spectra, electron microscopy images and molecular structures. These data types differ in format, scale and semantic meaning, making them difficult for text-only language models to handle together. MatVerse uses an independent encoder and unified decoder architecture: separate encoders extract features from each modality, while a shared decoder aligns them in a common semantic space. This enables different instruments, scales and data formats to be represented, associated and reasoned over within one model. A key claim of the project is that the system moves beyond knowledge retrieval toward mechanism-level inference. The team applied chain-of-thought reinforcement learning strengthened by scientific knowledge and molecular mechanisms, allowing the model to connect materials outcomes with underlying physical and chemical processes rather than only recalling published findings. The researchers also introduced the OmniMat agent framework, which supports cross-disciplinary literature search, structured hypothesis generation, mechanism-based candidate screening and optimization through expert and experimental feedback. To evaluate the model, the team developed OmniMatBench, a benchmark covering 19 secondary disciplines in materials science, and integrated it into the Sinan artificial intelligence evaluation system. This provides a standardized way to test capability limits in specialized research scenarios. The model’s engineering pipeline was completed on domestic computing platforms including Ascend and Sugon, covering data cleaning, architecture design, training, inference and safety alignment. MatVerse is being used in explorations of energy materials, low-dimensional materials, fibers, solid-state electrolytes, perovskites, two-dimensional materials and carbon nanotube fibers. In one validated case involving polymer thermoelectric batteries, the system extracted transferable mechanisms from 500 interdisciplinary papers, generated thousands of candidate hypotheses and converged on executable preparation routes. Real fabrication confirmed a device producing about 245 millivolts under a 10-kelvin temperature difference, with a normalized power density of 0.68 milliwatts per square meter per kelvin squared, suggesting possible human-body heat energy harvesting. Similar workflows have been applied to new elastomer materials for humanoid robot biomimetic muscles. The open-source MatVerse model family has recorded more than 300,000 downloads. The model is also expected to be integrated into the Shusheng Duan Yan scientific discovery platform, a full-stack research execution system spanning hypothesis generation and experimental validation. The platform has already been applied in life sciences, critical materials, semiconductors, fusion, quantum science and Earth-system modeling. Chen Xin, a principal investigator at Suzhou National Laboratory and deputy director of its artificial intelligence research department, presented the work at the event. The project team included graduate students and researchers from materials science, chemistry, physics, mathematics, computer science and artificial intelligence. Their papers received recognition as a 2025 outstanding paper from Cell Reports Physical Science and were selected for the journal’s list of most influential papers. The research has also been included as a case study in materials artificial intelligence textbooks.