Nitinol Interatomic Potential Using Moment Tensor Potentials in Machine Learning

In this study we developed an interatomic potential for the metal compound Nitinol using moment tensor potentials (MTP) by employing machine learning. The traditional way of determining interatomic forces is through quantum mechanics, which often requires a high computing cost and time. MTP has been shown to efficiently generate interatomic potentials for a wide variety of metallic systems. Although a Deep Learning potential for Nitinol has already been reported, it often requires high GPU and processing costs. Using MTP would allow for an affordable alternative potential that closely resembles the efficacy of Deep Learning Neural Network potential. The results presented were found using MTP codes run on a local AI workstation at Missouri State University (MSU). By optimizing the hyperparameters, errors in the interatomic potential were reduced dramatically approaching over 90 percent accuracy. We also demonstrated the benefit in utilizing the active learning algorithm to further enhance its accuracy. The computational works were performed at NERSC at National Lawrence Berkeley Laboratory and MSU's AI workstation. The support from NASA-Missouri Space Grant Consortium is acknowledged.