Physics-Informed Neural Approaches for Multiscale Molecular Modeling and Design

• Main Author: Qiao, Zhuoran
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2023
• Subjects: Artificial intelligence; Binding sites; Chemical elements; Chemistry; Computational chemistry; Deep learning; Energy; Geometry; Graph representations; Machine learning; Neural networks; Phase transitions; Physics; Success
• ISBN: 9798379854799
• DOI: 10.7907/48d1-ja21

Chemical processes in nature span multiple characteristic length and time scales, and the computational simulation for systems at the intersection of different scales is highly challenging with far-reaching implications for numerous scientific and industrial problems. To facilitate the computational modeling and design for large molecular systems and address the cost-resolution tradeoffs in conventional strategies, in this dissertation we introduce a series of physics-informed machine learning methods for the efficient computational modeling of chemical systems and the accurate prediction of their properties such as energetics, structures, and dynamics. In Chapters 2-3, we introduce a family of orbital-based geometric deep learning methods for the prediction of quantum chemical properties while adhering to the scaling and symmetry constraints of electronic structure theory. The presented methods achieve a chemical accuracy on community-wide benchmarks for molecular property prediction, and are shown to be transferable among diverse main-group molecular systems. In Chapter 4, we introduce a method for the prediction of protein-ligand complex structures based on a finite-time stochastic process parameterized by deep equivariant neural networks. The presented method achieves improved structure prediction accuracy against existing approaches, and is able to rapidly sample protein structures for folding landscapes that are modulated by inter-molecular interactions.