Part of the TPC Seminar Series

---

Abstract:

This talk will highlight some on-going work at the Michigan Institute for Computational Discovery & Engineering (MICDE) in building towards a future of AI-augmented science. The first part of the talk will cover MIST, a molecular foundation model trained on Polaris that has been scaled to 9B molecules to date, achieving SOTA performance on multiple benchmarks. Notably, our novel tokenizer, SMIRK, losslessly tokenizes OpenSMILES, enabling MIST to directly interact with the encoded elemental, electronic, nuclear, and geometric features. As we close in on our final production run on  50B molecules, we are exploring downstream applications of MIST and how MIST’s learned embedding space can be used for meaningful chemical exploration. Next, we discuss progress towards developing foundation models in a spatio-temporal physics setting.  We show that incorporating physical constraints into diffusion models helps improve performance in forward and inverse problems governed by partial differential equations. Equipping these models with spatial feature cross-attention and conditional encoding supports probabilistic generation of high-quality fields from sparse observations. With refined sensing representations and an unraveled temporal dimension, our method can handle arbitrary moving sensors and effectively generates spatio-temporal fields in a probabilistic fashion. I will conclude the talk by presenting a vision for an agentic framework that integrates and orchestrates domain-specific foundation models such as the above towards the goal of autonomous scientific exploration and discovery.

Biography:

Karthik Duraisamy is a Professor of Aerospace Engineering at the University of Michigan (U-M)where he also directs the Michigan Institute for Computational Discovery and Engineering (MICDE).He holds a PhD in Aerospace Engineering and a Masters in Applied Mathematics from the Universityof Maryland. His research interests span various aspects of computational science and AI includingdata-driven and reduced order modeling, statistical inference, numerical methods and Generative AIfor science. He is the PI of the U-M/Los Alamos Center on Advanced Computational Sciences. He isalso the founder and chief scientist of the silicon-valley-based startup Geminus.AI, which is focused onphysics-informed AI to accelerate autonomous industrial operations.