Abstract

Reactive machine-learned potentials have recently emerged as one of the most promising computational tools for materials chemistry, allowing us to apply statistical mechanical methods to more realistic and complex problems than ever before. Many of these problems require training models of exotic environments and non-equilibrium boundary conditions, where existing model architectures and training protocols may not be suitable. These limitations are only now beginning to be explored.

In this talk, I will show that the inherent localisation of most symmetry functions prevents standard machine-learned models from describing liquid-vapour interfaces correctly, since many properties of these interfaces are sensitive to unbalanced long-range forces such as electrostatic interactions. Using the perspective of local molecular field (LMF) theory due to Weeks and coworkers, I will show how neural network potentials attempt to represent these long-ranged forces and where they fail. The LMF approach suggests a simple modification to the neural network method that dramatically improves the description of polarisation-dependent properties. I will also discuss practical details of our neural network protocols, including refinements to the standard loss functions and our approach to data generation, both of which improve training efficiency and model generalisation.