(538a) Single-NN: A Modified Behler-Parinello Neural Network with a Single Neural Network

Machine learning (ML) has been increasingly applied in the field of computational chemistry to simulate the potential energy surface (PES) of chemical systems. Various methods including the Behler-Parinello Neural Network (BPNN), Crystal Graph Convolutional Neural Networks (CGCNN), SchNet, etc. have been developed for this purpose. In this work, we propose the SingleNN, which is a modified version of the BPNN. The modification is that instead of using one NN for each element, all elements share the same NN that has multiple outputs which correspond to the atomic energy for each of the elements. The SingleNN has some benefits; it is easier to implement and it reduces the number of parameters that must be trained. This sacrifices the flexibility of the model, but it also constrains different elements to share the same parameters such that the same interaction information is learned. We will demonstrate the method with Au-Pd nanoparticles and slabs and show that it can achieve training accuracy on par with the BPNN.