Affordable Quantum Chemistry via Data-Driven and Local Approximations to Non-Covalent Interactions

Thumbnail Image
Glick, Zachary Lee
Sherrill, C. David
Associated Organization(s)
Organizational Unit
Supplementary to
Quantum chemistry (QC) calculations can provide physically-rooted insight into intermolecular interactions. A quantitative understanding of these interactions, in turn, is of crucial importance for chemical problems like the modeling of protein-ligand interactions or molecular crystals and clusters. Unfortunately, the expensive computational cost of QC calculations prohibits their routine use in high-throughput computational workflows. The field of machine learning (ML) offers a potential workaround to this problem. Large amounts of quantum chemistry data can be generated upfront and used to parameterize models such as neural networks (NNs). The ML models can then be used to predict QC properties of new chemical systems, usually with a many order-of-magnitude reduction in computational cost. The development of such models is a rapidly evolving field, and numerous open questions exist about functional forms, dataset generation, accuracy, and generalizability. In this thesis, the development of NNs specific to the prediction of long-range, non-local intermolecular interactions--which existing models are not equipped to capture--is explored. Throughout the course of the chapters two through four, an equivariant atomic-pairwise neural network with a hybrid force field functional form referred to as AP-Net is developed. In the interest of the efficient generation of QC datasets, chapter 5 is concerned with the development and implementation of reduced-scaling dispersion algorithm. This algorithm allows for reference interactions energies to be generated at a reduced computational cost.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI