Due to its computational efficiency compared to other quantum mechanical methods, Density Functional Theory (DFT) is a popular tool in computational chemistry. Recently, machine learning (ML) has been explored as a tool to develop more accurate exchange-correlation (XC) functionals, but more work is required to design ML models which are transferable across chemical space and can be interpreted in the context of conventional functional design. To this end, we introduce two developments to design functionals that are transferable, obey exact theoretical constraints, and have separate exchange and correlation parts. First, we design a Gaussian Process-based exchange-only functional that obeys the uniform scaling rule and approximately matches the homogeneous electron gas limit. Second, we explore the use of the exchange energy density (both exact and ML) as a parameter for the correlation functional, effectively resulting in a more flexible local hybrid without gauge ambiguity issues. The accuracy of these approaches is competitive with semi-empirical functionals and recent ML models for atomization energies, ionization potentials, and barrier heights.

Design of transferable and interpretable machine learning exchange-correlation functionals

ACS Spring Meeting 2021

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