Refreshments are available starting at 10:30 a.m. The seminar will begin at 10:45 a.m.
Abstract
Scientific discovery has historically relied on empirical observation and theoretical derivation, but today, AI is fundamentally transforming this paradigm to accelerate the pace of discovery in the physical sciences. This talk explores the core concepts of AI for science, focusing on revolutionizing materials discovery through advanced machine learning architectures. First, Keqiang Yan will discuss the necessity of capturing the underlying physical and geometric principles of materials by ensuring invariance, equivariance, and completeness in representations through architectures like Matformer, ComFormer, and HIENet, alongside a brief discussion of GMTNet for high-order general tensorial property prediction. Second, bridging the gap between geometric structures and natural language processing, Yan will present Mat2Seq, a framework utilizing invariant tokenization to enable state-of-the-art generative capabilities for crystalline materials. Third, he will outline the frontier of autonomous scientific discovery by introducing MAPPS, a multi-agent system designed to unify planning and physical constraints to autonomously discover materials with targeted properties, and he will also highlight ongoing research dedicated to piezoelectric materials discovery. Altogether, these advancements chart a promising path toward highly autonomous, efficient, and physically grounded AI pipelines for next-generation materials design.
Speaker Biography
Keqiang Yan is a postdoctoral research associate in the Department of Computer Science at Princeton University. His research focuses on AI for science—specifically, developing foundation models, large language models, and 3D geometric modeling for molecules, materials, and proteins, alongside autonomous multi-agent systems for scientific discovery. Yan’s work has appeared at leading machine learning venues, including the Conference on Neural Information Processing Systems, the International Conference on Machine Learning, and the International Conference on Learning Representations, as well as in interdisciplinary journals such as Science Advances. Among his notable research achievements, his methods achieved the #1 ranking for formation energy prediction on the Matbench leaderboard upon publication, and he was a leading member of the 3rd place team in the Open Catalyst Challenge. Furthermore, he is a dedicated contributor to the open-source AI for science community, co-developing widely used resources such as DIG and AIRS, which have collectively garnered over 2,700 stars on GitHub. Yan is a recipient of the prestigious D.E. Shaw Research Doctoral & Postdoctoral Fellowship and the Graduate Research Excellence Award (one per year) from the Department of Computer Science and Engineering at Texas A&M University.