Deterministic Optimisation of Jastrow Factors

TL;DR

This paper introduces a deterministic method for optimizing Jastrow factors in quantum chemistry, reducing noise and improving energy accuracy compared to traditional stochastic approaches.

Contribution

It presents a novel deterministic approach to optimize Jastrow factors by minimizing the variance of the transcorrelated energy, enhancing reproducibility and accuracy.

Findings

Achieves low-variance Slater-Jastrow wavefunctions comparable to VMC optimization.

Produces lower energies than variance-optimized VMC.

Applicable to first row atoms and molecules with improved results.

Abstract

Highly flexible Jastrow factors have found significant use in stochastic electronic structure methods such as variational Monte Carlo (VMC) and diffusion Monte Carlo, as well as in quantum chemical transcorrelated (TC) approaches, which have recently seen great success in generating highly accurate electronic energies using moderately sized basis sets. In particular for the latter, the intrinsic noise in the Jastrow factor due to its optimisation by VMC can pose a problem, especially when targeting weak (non-covalent) interactions. In this paper, we propose a deterministic alternative to VMC Jastrow optimisation, based on minimising the "variance of the TC reference energy" in a standard basis set. Analytic expressions for the derivatives of the TC Hamiltonian matrix elements are derived and implemented. This approach can be used to optimise the parameters in the Jastrow functions, either from scratch or to refine an initial VMC-based guess, to produce noise-free Jastrows in a reproducible manner. Applied to the first row atoms and molecules, the results show that the method yields Slater-Jastrow wavefunctions whose variances are almost as low as those obtained from standard VMC variance optimisation, but whose energies are lower, and comparable to those obtained from energy-minimisation VMC. We propose that the method can be used both in the context of the transcorrelated method or in standard VMC as a new way to optimise Jastrow functions.