SingleNN: Modified Behler–Parrinello Neural Network with Shared Weights for Atomistic Simulations with Transferability
Mingjie Liu, John R. Kitchin
Abstract
In this article, we introduce the SingleNN, which is a modified version of the Behler–Parrinello Neural Network (BPNN) where the neural networks for the prediction of atomic energy for different elements are combined into a single network with shared weights. Using a data set containing Cu, Ge, Li, Mo, Ni, and Si, we demonstrate that SingleNN could achieve an accuracy that is on a par with BPNN for energy and force predictions. Furthermore, we demonstrate that SingleNN could learn a common transformation for the fingerprints of atoms to a latent space in which the atomic energies of the atoms are nearly linear. Using the common transformation, we could fit the data with new elements by changing only weights in the output layer in the neural network. In this way, with a moderate compromise in accuracy, we can speed up the training process significantly and potentially reduce the amount of training data needed.