Machine Learning for Atomistic Simulation II: Electronic Structure and Long-Range Charge Interactions

This talk will cover two recent developments in the Cider framework [1, 2, 3] for machine learning density functionals. First, we will extend Cider to learn the correlation functional in addition to exchange, which requires additional considerations for incorporating exact constraints and accounting for the use of approximate densities and reference energies during training. We will present molecular and solid-state benchmarks for the accuracy and computational efficiency of full exchange-correlation functionals trained with Cider. Second, we will introduce CiderPress, an open-source code package for evaluating Cider functionals in different DFT codes. We will provide an overview of the various density and orbital-dependent descriptors available in the code and explain how CiderPress can fit functionals to the energies, band gaps, ionization potentials, and electron affinities of both isolated and periodic systems.

[1] K. Bystrom and B. Kozinsky, J. Chem. Theory Comput. 2022, 18, 4, 2180–2192.

[2] K. Bystrom and B. Kozinsky, Phys. Rev. B 110, 075130 (2024).

[3] K. Bystrom, S. Falletta, and B. Kozinsky, J. Chem. Theory Comput. 2024, 20, 17, 7516–7532.