Machine learning foundation models are expanding the range of chemical systems we can simulate and problems we can pursue, but generating their training data comes at substantial computational cost. How much of that data do we actually need? My group develops intelligent learning algorithms and builds known physics into models so they can learn more from less data.
I will first introduce AdaptNTK, a learning algorithm that selects training data for machine learning models of atomistic systems based on what each new example would teach the model.
Shuwen Yue,
Cornell University
By accounting for the information gained from each selection, it avoids redundant data. On a reaction benchmark, it reduced the training data needed by more than half compared with a standard approach, without sacrificing accuracy. I will then discuss how we extend strategic learning beyond simulation to guide experimental design. We can also reduce what a model needs to learn by building known physics into it. I will discuss how we incorporate long-range electrostatics and the interactions governing electronic states into models for chemical systems. These approaches aim to improve data efficiency while capturing physical effects that simpler models miss.
Shuwen Yue is an Assistant Professor in the R. F. Smith School of Chemical and Biomolecular Engineering at Cornell University. Her group develops physics-aware machine learning and molecular simulation methods to understand and control chemistry in complex environments. She received her PhD in Chemical and Biological Engineering from Princeton University and completed postdoctoral research in Chemical Engineering at MIT. She is a Scialog Fellow and Awardee in Sustainable Minerals, Metals, and Materials, a recipient of Cornell Engineering’s SPROUT Award, ACS PRF Doctoral New Investigator, and a Digital Discovery Emerging Investigator. She also serves on the Early Career Board of the Journal of Chemical Theory and Computation and as Liaison Director of AIChE’s Computational Molecular Science and Engineering Forum.