|
Machine Learning Diffusion Monte Carlo Energy Densities
K. Ryczko, J.T. Krogel, I. Tamblyn Journal of Chemical Theory and Computation, 18, 12, 7695-7701 (2022) Diffusion Monte Carlo (DMC) is one of the most accurate methods for computing ground-state energies of molecules and materials, but its high computational cost limits routine application. In this work, we demonstrate that machine learning models trained on relatively small DMC datasets of approximately 60 calculations can accurately predict DMC energies. We employ two approaches: voxel deep neural networks (VDNNs) that predict DMC energy densities from Kohn-Sham DFT electron densities, and kernel ridge regression (KRR) that predicts atomic contributions to total DMC energies using atomic environment vectors. We find that KRR outperforms VDNNs by an order of magnitude on pristine graphene lattices, achieving mean absolute errors below chemical accuracy across all tested systems. The KRR models successfully predict energy barriers for Stone-Wales defects and generalize to three-dimensional materials such as liquid water, proving to be more accurate than Kohn-Sham DFT in all cases tested. |
