SUSTech Professor Yu Peiyuan Starts AI+Materials Venture: The Real Challenge Is Verification, Not Generation

Yu Peiyuan is currently a tenured associate professor (researcher) and doctoral supervisor in the Department of Chemistry at Southern University of Science and Technology, with long-term research in reaction mechanisms, quantum chemistry calculations, and materials design. He founded Sanxiang Epoch to build a verifiable materials model that addresses the gap between generation and validation in materials R&D.

Yu believes that the real difficulty in AI+ materials is not getting models to propose more plausible answers, but judging whether those answers can be synthesized, validated, and optimized. Successful experiments are valuable, but failed experiments should also feed back into the data and models, informing the next round of R&D.

Sanxiang Epoch's technical path revolves around 'verifiable reaction-native materials models' and is structured in three layers: React-Generator generates candidates, React-Refine performs high-precision computational screening, and React-Route/Condition plans reaction pathways and conditions. The third layer is where the team has invested the most and is their core moat.

The team has long-term expertise in reaction mechanisms, transition states, and quantum chemistry calculations. Two works published in Science demonstrate their capabilities in designing entirely new synthetic pathways and high-throughput chiral molecule synthesis. These strengths are being translated into React-Route/Condition, enabling the model to analyze 'how it should be done' and 'where it might get stuck.'

Yu emphasizes that validation should begin right after candidate generation, not just iterate based on final experimental results. He cites A-Lab as an example: A-Lab selected 57 theoretically stable target materials, successfully synthesized 36, failed to obtain target products for 17, and could not confirm 4 by XRD alone. Out of 353 experimental formulations tested, only about 30% yielded the target material.

Failure reasons include slow reaction kinetics, precursor volatility, material amorphization, and computational judgment biases. This shows that theoretical stability only means a material 'might exist'; whether it can actually be made depends on the reaction process, synthetic pathway, and experimental conditions.

Yu believes data is a core moat, but more data is not necessarily more valuable. The team aims to accumulate complete, traceable R&D paths, including both success and failure data, especially failure data that can help the model eliminate wrong directions earlier.

The team has accumulated millions of high-precision computational and generative data points, with key mechanisms and reaction pathways being cross-validated through collaborative experiments. React-DB aims to supplement the dynamic processes not covered by Materials Project, recording reaction pathways, experimental conditions, and failure information from raw materials to products.

Sanxiang Epoch has two business paths: one is joint R&D with enterprises, generating revenue from R&D service fees, milestone payments, and licensing of results; the other is proprietary materials IP development, such as liquid cooling materials and next-generation OLED emitters, capturing long-term value through patent licensing or joint industrialization.

The team is built around 'models that can compute, experiments that can be done, and industry that can use it.' Yu Peiyuan is responsible for overall strategy and mechanism calculations, while Professor Tan Bin serves as co-founder and chief scientist, overseeing synthetic route development and experimental validation. The AI and automation components are handled by the core technical team.