Force training neural network potential energy surface models

Abstract

Machine learned chemical potentials have shown great promise as alternatives to conventional computational chemistry methods to represent the potential energy of a given atomic or molecular system as a function of its geometry. However, such potentials are only as good as the data they are trained on, and building a comprehensive training set can be a costly process. Therefore, it is important to extract as much information from training data as possible without further increasing the computational cost. One way to accomplish this is by training on molecular forces in addition to energies. This allows for three additional labels per atom within the molecule. Here we develop a neural network potential energy surface for studying a hydrogen transfer reaction between two isomers of $C_5{H}_5$. We show that, for a much smaller training set, force training not only improves the accuracy of the model compared to only training on energies, but also provides more accurate and smoother first and second derivatives that are crucial to run dynamics and extract vibrational frequencies in the context of transition-state theory. We also demonstrate the importance of choosing the proper force to energy weight ratio for the loss function to minimize the model test error.