Extending the Reach of Quantum Monte Carlo Methods via Machine Learning

Quantum Monte Carlo methods are a promising suite of stochastic electronic structure methods that enable the high-accuracy modeling of strongly correlated molecules and materials at comparatively modest costs through the sampling of random numbers. Historically, these methods have excelled at computing energies, but have struggled to efficiently compute forces and scale to large system sizes that approach the thermodynamic limit. In this work, we illustrate how machine learning can address these critical shortcomings. We will first describe our recent endeavors employing active learning with AMPTorch to predict QMC-quality forces for the relaxation of molecular geometries and molecular dynamics simulations. We will illustrate how active learning is particularly crucial in the QMC context because of the lack of forces that can be leveraged for training. In the second portion of this talk, we will subsequently illustrate how Gaussian Process Regression (GPR) can be used to more accurately predict the energies of solids in the thermodynamic limit than conventional scaling equations. Altogether, this work demonstrates how machine learning can be used to extend the capabilities of stochastic methods, increasing their overall practicality and applicability.