Material properties prediction using machine learning-based ab initio calculations

We present the study of statistical mechanics properties of materials using machine learning-based models and developing first principles derived Hamiltonian to connect to Monte Carlo simulations. The evaluation of the system’s Hamiltonian at ab initio accuracy is a highly expensive step in the Monte Carlo method when sampling the phase space. The computational cost can be reduced when using machine learning-based models as surrogates for the ab initio calculations within the classical Monte Carlo sampling of the phase space. In this project, we have compared the performance results from using a linear mixing model and HydraGNN on different compositions of NiPt alloy. These two surrogate models are used to learn the interactions of their constituents. We present the predictive performance of these two surrogate models with respect to their complexity while avoiding the danger of overfitting the model. Our results show that the HydraGNN model attains superior predictive performance for magnetic alloy materials. We will also show that the models trained on smaller system sizes can be used to predict the properties of NiPt alloys of larger system sizes.

 This research used resources of the Oak Ridge Leadership Computing Facility, which is supported by the Office of Science of DOE under Contract No. DE-AC05-00OR22725.