Localized resolution of identity approach to the analytical gradients of random-phase approximation ground-state energy: algorithm and benchmarks

TL;DR

This paper introduces a localized resolution of identity method for calculating analytical gradients of RPA ground-state energies, enabling accurate molecular geometry optimizations and benchmarks with high precision.

Contribution

It presents a novel formalism and implementation for RPA energy gradients using LRI and density functional perturbation theory, allowing geometry relaxations at the RPA level.

Findings

High numerical precision in RPA gradient calculations.

RPA overestimates bond lengths systematically.

RPA energy hierarchy for water hexamers matches coupled cluster results.

Abstract

We develop and implement a formalism which enables calculating the analytical gradients of particle-hole random-phase approximation (RPA) ground-state energy with respect to the atomic positions within the atomic orbital basis set framework. Our approach is based on a localized resolution of identity (LRI) approximation for evaluating the two-electron Coulomb integrals and their derivatives, and the density functional perturbation theory for computing the first-order derivatives of the Kohn-Sham (KS) orbitals and orbital energies. Our implementation allows one to relax molecular structures at the RPA level using both Gaussian-type orbitals (GTOs) and numerical atomic orbitals (NAOs). Benchmark calculations show that our approach delivers high numerical precision compared to previous implementations. A careful assessment of the quality of RPA geometries for small molecules reveals that post-KS RPA systematically overestimates the bond lengths. We furthermore optimized the geometries of the four low-lying water hexamers -- cage, prism, cyclic and book isomers, and determined the energy hierarchy of these four isomers using RPA. The obtained RPA energy ordering is in good agreement with that yielded by the coupled cluster method with single, double and perturbative triple excitations, despite that the dissociation energies themselves are appreciably underestimated. The underestimation of the dissociation energies by RPA is well corrected by the renormalized single excitation correction.