Accurate neural-network-based fitting of full-dimensional N₂ –Ar and N₂–CH₄ two-body potential energy surfaces aimed at spectral simulations

Abstract

We describe the development of machine-learned (ML) potentials for flexible, weakly interacting monomers. A recently suggested permutationally invariant polynomial neural network (PIP-NN) approach is utilised to represent the full-dimensional two-body component of the molecular pair energy. To ensure the asymptotic zero-interaction limit, a tailored subset of the full invariant polynomial basis set is utilised, and their variables are modified to achieve a better fit of the correct asymptotic behaviour at a long range. This new technique is used to build full-dimensional potentials for the two-body N${}_2$–Ar and N${}_2$–CH${}_4$ interactions by fitting databases of ab initio energies calculated at the coupled-cluster level of theory. The second virial coefficient, fully accounting for molecular flexibility, is then calculated within the classical framework using the obtained PIP-NN potential surfaces. A trajectory-based simulation of the N${}_2$–Ar collision-induced absorption is conducted, covering both the far- and mid-infrared ranges.

GRAPHICAL ABSTRACT

KEYWORDS:

• Intermolecular interaction
• machine learning
• collision processes

Acknowledgments

We were able to carry out ab initio calculations and train ML PESs thanks to the Smithsonian Institution's allocation of computing time on the HPC and GPU facilities of the Smithsonian Institution High Performance Cluster [73]. Additionally, I am thankful to Daniil Chistikov and Andrey Vigasin for their insightful comments and helpful discussions.

Data availability

Supplementary material contains full-dimensional two-body PIP-NN model implementations in C++ for the N${}_2$–Ar and N${}_2$–CH${}_4$ systems. We also provide a PIP-NN model implementation for ethanol that was fitted to the training and validation budget of 50,000 energies and forces taken from the MD17 database. The interaction energy data sets for N${}_2$–Ar and N${}_2$–CH${}_4$, as well as code for training the PIP-NN models, can be found at https://github.com/artfin/PES-Fitting-MSA.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 The data set was obtained from http://quantum-machine.org/gdml/##datasets.

Additional information

Funding

The research was supported by the non-commercial Foundation for the Advancement of Science and Education INTELLECT. I acknowledge the support by Russian Science Foundation Grant No. RSCF 22-17-00041 in developing machine-learned two-body PES for the N${}_2$–CH${}_4$ system.