Electron‐Passing Neural Networks

Abstract

While neural-network force fields based on the Behler-Parrinello (BP) architecture already make the construction of highly accurate and fast-to-evaluate potential energy hypersurface representations possible, there still remain open challenges. Among these is the inclusion of long-range interactions and the modelling of forces due to externally applied electric fields. We present an investigation of electron-passing neural networks (EPNN) as an option to tackle these challenges. Emphasis is put on finding physically well-defined target quantities for prediction.