Van-der-Waals interactions in neural-network force fields

Abstract

In recent years, new methods based on machine-learning techniques and neural-networks have been developed to calculate the thermodynamic and spectroscopic properties of various systems with high precision and low computational costs. However, most current neural-network force fields have a significant limitation: they only take into account local, short-range interactions between atoms and molecules, while largely ignoring the longrange interactions that may strongly influence the properties of materials. This limitation may have significant consequences for predicting the physical properties of materials that are primarily determined by long-range interactions. This master thesis discusses the extent to which long-range interactions are taken into account in the current neuralnetwork force field NeuralIL, and how this limitation affects the accuracy of predicted density using molecular dynamics simulations for water as a case study. Moreover a new implementation of the DFT-D3 method for calculating dispersion interactions using the high-performance machine-learning framework JAX will be presented. Finally, the suitability of vdW-DF exchange-correlation functionals to correct for those effects in condensed-phase water systems will be analysed and compared to DFT calculations with RPBE functional with and without the DFT-D3 dispersion correction.