Simulations of water and hydrophobic hydration using a neural network potential

TL;DR

This paper demonstrates that a neural network potential trained solely on quantum data can accurately simulate water and hydrophobic hydration, bridging high-level ab initio methods with large-scale molecular simulations.

Contribution

The study introduces ANI-1ccx, a neural network potential that accurately models water and hydrophobic hydration without using liquid phase data for training.

Findings

Neural network potential matches ab initio accuracy for water simulations

Model successfully simulates aqueous methane solvation

No liquid phase data needed for training

Abstract

Using a neural network potential (ANI-1ccx) generated from quantum data on a large data set of molecules and pairs of molecules, isothermal, constant volume simulations demonstrate that the model can be as accurate as ab initio molecular dynamics for simulations of pure liquid water and the aqueous solvation of a methane molecule. No theoretical or experimental data for the liquid phase is used to train the model, suggesting that the ANI-1ccx approach is an effective method to link high level ab initio methods to potentials for large scale simulations.