This was part of Materials Informatics: Tutorials and Hands-On

Generative Models for Molecules and Materials: Fundamentals, Opportunities, and Open Problems

Grant Rotskoff, Stanford University

Thursday, March 14, 2024

Abstract: This talk will give a detailed introduction to generative models with an emphasis on design principles that apply to molecular systems. We will first discuss several relevant paradigms for building generative models, including score-based diffusion and continuous normalizing flows. We will then detail how both intramolecular and intermolecular constraints inform the architectures appropriate for modeling nanoscale materials. Finally, we will come to the crucial question of how to build scalable training strategies for systems where thermal fluctuations cannot be neglected. The talk will end with a discussion of the most pressing open problems for generative "AI" in the physical sciences.