EAGER: Density functionals from Machine Learning

University Of California-Irvine, Irvine CA

Investigators

Abstract

Kieron Burke of the University of California, Irvine, is supported by an Eager award from the Chemical Theory, Models and Computational Methods program in the Chemistry Division. The award is cofunded by the Condensed Matter and Materials Theory program in the Division of Materials Research and the Computational and Data Enabled Science Program in the Division of Mathematical Science. Density functional theory (DFT) is at the heart of modern electronic structure calculations, which play an ever increasing role in chemical and material design. Present-day approximations are created by an unholy alliance of inspiration and pragmatism. Progress in their improvement is slow and unsystematic. Burke and his colleagues are applying Machine Learning (ML) to the problem of approximating density functionals. ML provides a totally new way to approximate functionals that takes maximum advantage of DFT formalism. They have demonstrated that chemical accuracy on self-consistent densities can be reached for a simple model case, and a measure of the reliability of the approximation can be given. The goal of the EAGER grant is to develop these methods in application to DFT, overcoming any challenges, and reaching the real world of electronic structure calculations as quickly as possible, by approaching the problem in a well-defined series of small steps. This project creates an entirely new subfield of theory/computation, in which machine learning (ML), a branch of computer science, is applied to electronic structure problems, a branch of theoretical physics and chemistry that allows prediction of new molecules and materials by solving the equations of quantum mechanics. New algorithms, at the cutting edge of ML research, are being developed in order to apply ML to find density functionals, needed for solving electronic structure. This is a true synergy of physical and computer sciences. Success of this proposal would revolutionize materials design, allowing millions of atoms to be treated instead of hundreds by present methods. This would truly transform predictive capability in a broad range of scientific and technological problems, from biomolecular liquid simulations to crack propagation in materials.