Extension of selected configuration interaction for transcorrelated methods

TL;DR

This paper extends selected configuration interaction algorithms to the transcorrelated framework, demonstrating faster convergence and emphasizing the importance of non-Hermitian considerations for improved computational efficiency in atomic and molecular systems.

Contribution

It introduces a general framework for SCI in the transcorrelated method, accounting for non-Hermitian effects and exploring different selection criteria and perturbative corrections.

Findings

Non-Hermitian character is crucial for fast convergence.

Selection criteria significantly impact convergence rates.

SCI in TC converges faster than traditional SCI.

Abstract

In this work we present an extension of the popular selected configuration interaction (SCI) algorithms to the Transcorrelated (TC) framework. Although we used in this work the recently introduced one-parameter correlation factor [E. Giner, J. Chem. Phys., 154, 084119 (2021)], the theory presented here is valid for any correlation factor. Thanks to the formalization of the non Hermitian TC eigenvalue problem as a search of stationary points for a specific functional depending both on left-and right-functions, we obtain a general framework allowing different choices for both the selection criterion in SCI and the second order perturbative correction to the energy. After numerical investigations on different second-row atomic and molecular systems in increasingly large basis sets, we found that taking into account the non Hermitian character of the TC Hamiltonian in the selection criterion is mandatory to obtain a fast convergence of the TC energy. Also, selection criteria based on either the first order coefficient or the second order energy lead to significantly different convergence rates, which is typically not the case in the usual Hermitian SCI. Regarding the convergence of the total second order perturbation energy, we find that the quality of the left-function used in the equations strongly affects the quality of the results. Within the near-optimal algorithm proposed here we find that the SCI expansion in the TC framework converges faster than the usual SCI both in terms of basis set and number of Slater determinants.