A Fourth-Generation High-Dimensional Neural Network Potential with Accurate Electrostatics Including Non-local Charge Transfer*

Machine learning potentials have become a widely used tool for atomistic simulations.
However, most of the currently used methods predict the total energy based on local descriptors and are therefore unable to correctly describe effects caused by global changes in the electronic structure, such as long range charge transfer or different charge states.
Our fourth-generation high-dimensional neural network potential overcomes these limitations by introducing accurate atomic charges, obtained from a charge equilibration scheme based on environment dependent electronegativities, into the description of short ranged interactions.
The methods significantly improved description of the potential energy surface substantially extends the applicability of modern machine learning potentials.
This is demonstrated on several example systems which current methods fail to describe correctly.

Ko, T. W., Finkler, J. A., Goedecker, S., & Behler, J. (2020). A fourth-generation high-dimensional neural network potential with accurate electrostatics including non-local charge transfer. arXiv preprint arXiv:2009.06484.