Machine learning configuration interaction for ab initio potential energy curves

TL;DR

This paper advances machine learning configuration interaction (MLCI) methods to efficiently compute accurate ab initio potential energy curves, incorporating neural networks for configuration selection, duplicate removal, and spin state preservation, demonstrated on molecules like N₂, H₂O, and CO.

Contribution

The authors develop an improved MLCI approach that uses neural networks as hash functions and includes configuration state functions for spin purity, enhancing efficiency and transferability across geometries.

Findings

MLCI achieves high accuracy in potential energy curves for N₂, H₂O, and CO.

The method outperforms stochastic configuration selection in accuracy and efficiency.

Potential energy curves are computed with significantly less processor time.

Abstract

The concept of machine learning configuration interaction (MLCI) [J. Chem. Theory Comput. 2018, 14, 5739], where an artificial neural network (ANN) learns on the fly to select important configurations, is further developed so that accurate ab initio potential energy curves can be efficiently calculated. This development includes employing the artificial neural network also as a hash function for the efficient deletion of duplicates on the fly so that the singles and doubles space does not need to be stored and this barrier to scalability is removed. In addition configuration state functions are introduced into the approach so that pure spin states are guaranteed, and the transferability of data between geometries is exploited. This improved approach is demonstrated on potential energy curves for the nitrogen molecule, water, and carbon monoxide. The results are compared with full configuration interaction values, when available, and different transfer protocols are investigated. It is shown that, for all of the considered systems, accurate potential energy curves can now be efficiently computed with MLCI. For the potential curves of N$_{2}$ and CO, MLCI can achieve lower errors than stochastically selecting configurations while also using substantially less processor hours.