Neural Network Potentials for Chemical Reactivity

Christopher Stein, University of Duisburg-Essen, Germany

Recent years have seen a tremendous success of neural network potentials (NNPs) in materials simulation. These data-driven approaches easily outperform traditional quantum-chemical methods in terms of computational cost, while frequently maintaining the accuracy of the underlying parent method. The simulation of chemical reactivity – rare events with corresponding sharp features on the potential energy surface (PES) – is, however, a challenge for most NNPs.
In this talk, I will present NewtonNet, a NNP that was specifically designed to meet these challenges in the context of hydrogen combustion. I will discuss the design principles that enable NewtonNet to be more data efficient than competing NNPs and show how we intend to significantly reduce the required learning data with an active learning approach based on a dimensionality reduction of the relevant parts of the PES. Finally, I will advocate for other performance measures than standard learning curves to judge an NNPs ability to accurately describe chemical reactivity.

back to MCIC 2022