Many-Body Permutationally Invariant Polynomial Neural Network Potential Energy Surface for N₄.

Area of Science:

• Computational Chemistry
• Chemical Physics
• Materials Science

Background:

• Accurate potential energy surfaces (PES) are crucial for modeling chemical dynamics in high-energy environments such as shock waves and plasmas.
• Existing methods for fitting PES struggle with complex systems like N4, necessitating improved computational approaches.

Purpose of the Study:

• To develop and validate a neural network (NN) based method for fitting the many-body (MB) component of the N4 ground-state PES.
• To improve the accuracy and applicability of PES for high-energy molecular collision dynamics.

Main Methods:

• Employed neural networks (NNs) with permutationally invariant polynomials (PIPs) to fit the MB component of the N4 PES.
• Extended the fitting dataset with 4859 new CASPT2 points and 13 new CCSD(T) points, totaling 21,406 points.
• Implemented strategies for complete geometry domain coverage, including trajectory calculations and sparse data region searches.

Main Results:

• Achieved a highly accurate MB-PIP-NN fit for the N4 system, outperforming conventional least-squares methods by a factor of 3 in accuracy.
• The NN fit demonstrated impressive performance on a dataset with a wide energy range and rugged surface regions due to locally avoided crossings.
• While NN force calculations were an order of magnitude slower, the achieved accuracy is highly significant for complex systems.

Conclusions:

• The NN-based approach provides a powerful tool for generating accurate PES for challenging chemical systems.
• This method holds promise for accurately modeling chemical dynamics in extreme conditions where traditional methods fall short.
• The developed PES is expected to significantly advance research in shock wave and plasma chemistry.