Scalable Quantum Monte Carlo with Direct-Product Trial Wave Functions

TL;DR

This paper introduces a scalable quantum Monte Carlo method using direct-product trial wave functions to significantly reduce computational costs for molecules with decomposable active spaces, while maintaining accuracy.

Contribution

The authors propose a novel approach employing direct-product multi-Slater determinant trials for MSD-AFQMC, enabling efficient calculations for certain strongly correlated systems.

Findings

Reduced computational cost by up to 36 times for specific molecules.

Maintained accuracy for systems with decomposable active spaces.

Efficiency decreases in strongly coupled active subspaces.

Abstract

The computational demand posed by applying multi-Slater determinant trials in phaseless auxiliary-field quantum Monte Carlo methods (MSD-AFQMC) is particularly significant for molecules exhibiting strong correlations. Here, we propose using direct-product wave functions as trials for MSD-AFQMC, aiming to reduce computational overhead by leveraging the compactness of multi-Slater determinant trials in direct-product form (DP-MSD). This efficiency arises when the active space can be divided into non-coupling subspaces, a condition we term ``decomposable active space''. By employing localized-active space self-consistent field wave functions as an example of such trials, we demonstrate our proposed approach across a range of molecular systems, each exhibiting varying degrees of complexity in their electronic structures. Our findings indicate that the compact DP-MSD trials can reduce computational costs substantially, by up to 36 times for the C2H6N4 molecule where the two double bonds between nitrogen N=N are clearly separated by a C-C single bond, while maintaining accuracy when active spaces are decomposable. In the case of larger systems such as the benzene dimer, characterized by weak coupling between the two monomers, we observed a decrease in computational cost compared to using a complete active space trial, yet we retained the same level of accuracy. However, for systems where these active subspaces strongly couple, a scenario we refer to as "strong subspace coupling", the method's accuracy decreases compared to that achieved with a complete active space approach. We anticipate that our method will be beneficial for systems with non-coupling to weakly-coupling subspaces that require local multireference treatments.