Ming Hu1
University of South Carolina1
Ming Hu1
University of South Carolina1
The compromise between density functional theory (DFT) accuracy and molecular dynamics (MD) speed from machine learned potentials (MLPs) permits the simulation of larger and longer scales that were previously difficult to realize. A recently developed potential, namely the Deep Potential (DeepPot), is a novel neural network-based atomic potential with excellent DFT-level energy accuracy. However, the scalability of DeepPot with number of elements (M) is non-linear, where N element-specific networks require the allocation of N x M inputs during training. By construction, using a dataset containing many elements slows down training and reduces the quality of the prediction. Here, we present a modified deep neural network potential, named Elemental DeepPot, which introduces species-specific descriptors. For demonstration, we apply our modified deep neural network to a large dataset containing a mixture of ~12,000 metallic and semiconducting materials, totaling >50 unique elements for training. The results show Elemental DeepPot can capture both atomic structure and species with significant improvement to training speed and accuracy. We finally obtained a few thousand phonon dispersions and lattice thermal conductivities with accuracy comparable to full DFT calculations.