First-principles quantum mechanical modeling has become the primary workhorse for the high-throughput design of multi-fragment systems with compelling electronic and optical properties. However, it is infeasible when the target systems are too large due to the difficulty in balancing efficiency and accuracy. Many-body (fragment) expansion theory (MBET) has been developed to resolve this issue, but it still suffers from expensive first-principles evaluations of many-fragment energies. Our study replaced first-principles approaches for many-fragment energies with machine learning algorithms using frontier graph convolutional networks (GCNs). Operationally, we established quantitative relationships between effortless molecular descriptors and targeted many-fragment energies for multi-fragment systems with moderate or strong many-fragment interactions. As the initial stage of the study, we focused on van der Waals or hydrogen-bonded molecular aggregates, including water cluster, phenol cluster, and water-phenol mixture. Given sufficient configurational diversity in the training set, we successfully reproduced first-principles two-fragment and three-fragment energies to the order of 10<sup>-5</sup> kcal/mol but at a fractional computational cost. Our results indicated that GCN-aided MBET provides a promising alternative scheme to determine many-fragment energies and eventually desired electronic and optical properties for multi-fragment system energies.

Accelerating many-body expansion theory through GCN-based deep learning approaches

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.