NN-VRCTST: Neural Network Potentials Meet Variable Reaction Coordinate Transition State Theory for the Rate Constant Determination of Barrierless Reactions

TL;DR

This paper introduces a new method called NN-VRCTST that uses neural networks to speed up the calculation of reaction rates for barrierless chemical reactions.

Contribution

The novel approach combines neural network potentials with variable reaction coordinate transition state theory to reduce computational costs.

Findings

NN-VRCTST achieves accuracy within 20% of traditional VRC-TST simulations.

The method reduces the number of single-point energy evaluations by at least a factor of 4.

ANN training and VRC-TST calculations can be decoupled for better optimization and data quality inspection.

Abstract

The determination of rate constants for barrierless reactions
poses
severe problems from a theoretical perspective. The main challenges
concern the proper description of the electronic structure of the
reacting system, which may have multireference character, the anharmonicity
of the relative motions of the fragments, and the proper definition
of the reaction coordinate. The literature state of the art in the
context of transition state theory is its variable reaction coordinate
implementation (VRC-TST), which overcomes these difficulties in determining
the number of transition state ro-vibrational states through a Monte
Carlo sampling of the potential energy surface (PES) defined by the
relative orientation of the two fragments. Although approaching the
accuracy of experiments, VRC-TST requires tens of thousands of single-point
energy (SPE) evaluations, thus being computationally demanding. The
approach developed in this work, named NN-VRCTST, aims at fitting
the PES with physics-inspired artificial neural network (ANN) models
to be used as surrogate potentials in VRC-TST simulations. The ANN
efficacy is evaluated in the computation of high-pressure limit rate
constants for gas-phase barrierless reactions and validated over state-of-the-art
VRC-TST simulations. It is shown that the NN-VRCTST tool reaches an
accuracy within 20% with respect to VRC-TST simulations performed
by using traditional approaches. While lowering the number of SPE
needed by at least a factor of 4, the computational framework devised
here allows one to decouple ANN training and VRC-TST calculations,
enabling the optimization of the SPE evaluations as well as the quality
inspection of the employed data points. We believe that the NN-VRCTST
approach has the potential to evolve into a robust and computationally
efficient framework for performing VRC-TST calculations for barrierless
reactions.

Figures