Learning to Model Atoms Across Scales

Abstract

The understanding of atoms and how they interact forms the foundation of modern natural science, as well as material and drug discovery efforts. Computational chemistry methods such as density functional theory and molecular dynamics simulation can offer an unparalleled spatiotemporal resolution for observing microscopic mechanisms and predicting macroscopic phenomena. However, many natural processes are extremely complex, requiring highly accurate modeling of many atoms for a considerable period to study. Computational chemistry methods may not be accurate or efficient enough, limiting the applicable domains and scales. Furthermore, discovering new materials and drugs requires novel candidate atomistic structures, which are conventionally based on heuristic or exhaustive search methods. This thesis presents machine learning methods for modeling atoms for tasks across different scales. First, we propose machine learning force fields that can decompose molecular interactions into fast and slow components, and then accelerate molecular simulations through multiscale integration. Second, we propose an end-to-end workflow for learning time-integrated coarse-grained molecular dynamics using multi-scale graph neural networks. Third, we propose diffusion models designed for periodic material structures that can enable the discovery of novel stable materials as well as material inverse design given a target property. The material diffusion model can be further extended to complex metal-organic frameworks with a multi-scale modeling approach.