Sampling rare events using nanostructures for universal Pt neural network potential

Abstract

The density functional theory (DFT) data-driven approach to generating potential energy surfaces using machine learning has been proven to quickly and accurately predict the molecular and crystal structures of various elements. However, training databases consisting of hundreds of well-known symmetric structures have shown fatal weaknesses in calculating amorphous or nano-scale structures. Ab-initio molecular dynamics (AIMD) simulations create a training set that compensates for these shortcomings, but there are still many rare event structures. Here we introduce a new method to easily enlarge the data diversity and dramatically reduce data points based on the highly defected nano structures for universal machine learned potential. Our potential applies to bulk and nano systems and has been shown to high accuracy and computational efficiency while requiring minimal DFT training data. The developed potential is expected to help observation of structural changes in the Pt-based nano-catalysts that have been difficult to simulate at the DFT-level.