A neural network-based four-body potential energy surface for parahydrogen

TL;DR

This paper develops a neural network-based four-body potential energy surface for parahydrogen using high-level ab initio calculations, enabling more accurate modeling of interactions in para-H$_2$ clusters.

Contribution

It introduces a novel neural network approach to construct a four-body PES for para-H$_2$, incorporating rescaling and empirical adjustments for improved accuracy.

Findings

The PES accurately captures four-body interactions at various distances.

The model combines ab initio data with empirical dispersion corrections.

Short-range interactions are predominantly repulsive.

Abstract

We present an isotropic ab initio (para-H$_2$)$_4$ four-body interaction potential energy surface (PES). The electronic structure calculations are performed at the correlated coupled-cluster theory level, with single, double, and perturbative triple excitations. They use an atom-centred augmented correlation-consistent double zeta basis set, supplemented by a $(3s3p2d)$ midbond function. We use a multilayer perceptron to construct the PES. We apply a rescaling transformation to the output energies during training to improve the prediction of weaker energies in the sample data. At long distances, the interaction energies are adjusted to match the empirically-derived four-body dispersion interaction. The four-body interaction energy at short intermolecular separations is net repulsive. The use of this four-body PES, in combination with a first principles pair potential for para-H$_2$ [J. Chem. Phys. 119, 12551 (2015)], and an isotropic ab initio three-body potential for para-H$_2$ [J. Chem. Phys. 156, 044301 (2022)], is expected to provide closer agreement with experimental results.