Towards Low-Cost Electronic Structure: Machine-Learning Strategies for Extending Fock-Corrected Density Functional Theory

Disciplines:

Chemistry, Computer Science

Mentor:

Thomas Miller,
Professor of Chemistry, (CCE),
tfm@caltech.edu

Background:

There is an ever-growing demand for inexpensive and accurate density functional theory (DFT) calculations in the study of a variety of complex chemical systems. New methods are needed to improve the balance between the cost of a DFT calculation and its accuracy. The Miller group has recently developed the Fock-corrected density-functional theory, a semi-empirical DFT method, to attain improvement in accuracy without losing the affordable cost of using minimal-basis DFT [J. Chem. Theory Comput., 12, 5811 (2016).]. However, the method is currently limited to a small number of elements (i.e., carbon and hydrogen). Working with a summer researcher, we propose to employ neural networks and other machine learning strategies to further test the FC-DFT approach and to efficiently parameterize FC-DFT for a greater range of elements.

Description:

1. Become familiar with mean-field electronic structure methods, such as DFT 2. Become familiar with the use of neural-network machine-learning algorithms 3. Write code for FC-DFT parameterization using machine-learning algorithms