Density Functional Theory (DFT) is the most widely used method in computational physics and chemistry. However, the accuracy of existing semi-local functionals is inadequate for many systems, leading to charge delocalization, self-interaction errors and inconsistent treatment of static correlation.

Using Gaussian process regression in combination with exact constraints and efficient descriptors, we are developing Machine Learning models to approximate the universal exchange-correlation (XC) functional of DFT. Our goal is to obtain a transferable functional capable of computing ground-state properties of systems that are too complex for conventional XC functionals and too large for advanced quantum chemistry methods.

 Our initial work developing the CIDER Exchange functional is available on arXiv at https://arxiv.org/abs/2109.02788, and the code used in that project is available on our Github page at https://github.com/mir-group/CiderPress.