This paper is part of a joint effort, led by Viktor Zaverkin, to facilitate the efficient simulation of molecular dynamics at a quantum mechanical level of accuracy. Since the accurate computation of forces between atoms using numerical methods is slow, a neural network (NN) surrogate model is employed. Once trained on data from the numerical methods, the surrogate model can predict forces at similar accuracy, but much more efficiently. One such surrogate model, called Gaussian Moments Neural Network (GM-NN), has been proposed by Zaverkin [1]. The awarded paper [2] improves the training speed of GM-NN by an order of magnitude while reaching equal or better accuracy. This makes GM-NN suitable for use in active learning methods, which repeatedly retrain the NN while using it to select the most interesting data. In follow-up work, the authors then study various active learning strategies with GM-NN [3].
[1] Viktor Zaverkin and Johannes Kästner, Gaussian moments as physically inspired molecular descriptors for accurate and scalable machine learning potentials, Journal of Chemical Theory and Computation, 2020
[2] Viktor Zaverkin, David Holzmüller, Ingo Steinwart, and Johannes Kästner, Fast and Sample-Efficient Interatomic Neural Network Potentials for Molecules and Materials Based on Gaussian Moments, Journal of Chemical Theory and Computation, 2021
[3] Viktor Zaverkin, David Holzmüller, Ingo Steinwart, and Johannes Kästner, Exploring chemical and conformational spaces by batch mode deep active learning, Digital Discovery, 2022
