A natively multimodal materials-science foundation model built by Suzhou National Laboratory with Shanghai AI Lab and Shanghai Jiao Tong University, presented at the China Association for Science and Technology's "新天工开物" achievements session on 2026-09-30 and written up by Shanghai AI Lab on 2026-10-28. Architecture as described by the team: independent per-modality encoders plus a unified decoder that aligns the encodings in one shared semantic space, so XRD spectra, electron- microscopy images and molecular-structure graphs are reasoned over alongside natural-language questions rather than being flattened into text. Post-training uses chain-of-thought reinforcement learning "dual-enhanced" with domain knowledge and physico-chemical mechanism, which the team frames as the move from knowledge retrieval to mechanism reasoning. Training and inference ran on domestic accelerators (Ascend and Sugon).

Around the model the team ships OmniMat, an agent framework closing a cross-disciplinary loop of inspiration retrieval → structured hypothesis generation → mechanism-level candidate screening → expert and experimental feedback, and OmniMatBench (arXiv 2605.29833), a human-calibrated multimodal benchmark of 3,171 expert-curated QA and calculation problems across all 19 materials-science subfields — a balanced 1,000-problem subset is public on HuggingFace under CC BY-NC 4.0 — now integrated into the lab's OpenCompass (司南) evaluation system. A reported application is a polymer thermal battery whose preparation route the system derived from ~500 cross-disciplinary papers and thousands of generated hypotheses, measured after real fabrication at ~245 mV across a 10 K gradient and a normalized power density of 0.68 mW·m-2K-2; the same workflow is reported for an elastomer used in humanoid-robot artificial muscle.

Self-reported figures to read as such: the 材华 series has "over 300,000 cumulative downloads since open-sourcing" per the lab post, and academician Sun Baode called it at the launch the world's first 10-billion-parameter-class (百亿参数) and first natively multimodal materials model. Neither the announcement nor the launch coverage names a weights repository, an exact parameter count, a base model or a license, and we could not locate the checkpoints on HuggingFace or ModelScope, so no scale fields are recorded here; the model is to be integrated into the lab's 书生·端砚 (Sage) scientific-discovery platform.

sciencemultimodalmaterialsagents

Related