QMCTorch: Molecular Wavefunctions with Neural Components for Energy and Force Calculations

Abstract

In this paper, we present results obtained using QMCTorch, a modular framework for real-space Quantum Monte Carlo (QMC) simulations of small molecular systems. Built on the popular deep learning library PyTorch, QMCTorch is GPU-native and enables the integration of machine learning-inspired components into the wave function ansatz, such as neural network backflow transformations and Jastrow factors, while leveraging efficient optimization algorithms. QMCTorch interfaces with two widely used quantum chemistry packages - PySCF and ADF - which provide initial values for the atomic orbital exponents and molecular orbital coefficients. In this study, we present wavefunction optimizations for four molecules: $H_2$, $LiH$, $Li_2$, and $CO$, using various wavefunction ansätze. We also compute their dissociation energy curves and the corresponding interatomic forces along these curves. Our results show good agreement with baseline calculations, recovering a significant portion of the correlation energy. QMCTorch provides a modular and extendable platform for rapidly prototyping new wavefunction ansätze, evaluating their performance, and analyzing optimization outcomes.