Towards chemical foundation models for digital prediction of experimental measurements

E. Annevelink
Physics Inverted Materials,
United States

Keywords: digital materials, modeling, simulation


To realize digital materials development, accurate and generalizable models are needed to describe the diverse chemical and multiscale interactions that govern materials manufacturing and operation. While foundation models provide generalized descriptions for making predictions across diverse inputs, they need to be trained on large datasets spanning their hypothesis space. Although large density functional theory datasets cover much of the periodic table, they often only sparsely cover the chemical space, while larger length scale datasets are even more sparse. Data, therefore, needs to be identified and accurately generated in order to improve models across both chemical and length-scale complexity. Here we discuss our approach to realizing multi-scale chemical foundation models based on our data-efficient active learning methodology for producing rich datasets and developing multiscale chemical foundation models.