Machine learning potentials provide a method for accurate ground state energy assessment at affordable computational cost. Deep neural networks using symmetry functions for atomic representation have demonstrated capability in capturing both intra- and intermolecular interactions. In this work, we produce machine learning potentials for CO2, O2, N2, and NO, with the goal of modeling reactive equilibria between the latter three species. Atomic and molecular potential energy surfaces are sampled via molecular dynamics simulations under various conditions, and the ground state energies of extracted geometries are evaluated using a dispersion-corrected DFT functional. Neural networks are then trained on datasets of these energies, producing machine learning potential models capable of capturing bother intra- and intermolecular interactions. These models are then retrained on high quality energies evaluated using CCSD(T) and CASSCF in order to produce wavefunction level accuracy models, and the resulting potentials are then transferable to, and affordable for, simulation of ensembles of molecules. These models are then employed in Monte Carlo simulations to predict vapor-liquid equilibria, reactive equilibria, and second virial coefficients. Current results include models for CO2 and N2 that reproduce DFT assessment of disperse intermolecular interactions to within one tenth of a kilocalorie per mole. Future work includes producing potentials for open shell species, retraining potentials with wavefunction data, and including electrostatics.

Machine learning potentials for modeling reactive equilibria between organic species

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.