Title:Statistical Mechanics of Materials using First Principles Calculations and Machine Learning

Abstract:We will describe our developments leading to an automated workflow approach that will utilize machine learning methods to combine first principles density functional theory (DFT) calculations with classical Monte-Carlo (MC) simulations. This method allows the investigation of the statistical mechanics and the finite temperature behavior of real materials with complex compositions and interactions of their constituents. The investigation of finite temperature properties using MC methods requires a large number of evaluations of the system's Hamiltonian to sample the phase space needed to obtain physical observables as function of temperature. DFT calculations can provide accurate evaluations of the energies, but they are too computationally expensive for routine simulations. We demonstrate a solution to this problem that harnesses the computational power of large massively parallel computers by combining classical MC calculations with our first principles multiple scattering electronic structure code (LSMS), which shows exceptional scaling on current HPC architectures such as Summit, by employing Machine Learning (ML) techniques. This workflow that can consider both classical interaction models and artificial neural network based models, allows us to investigate alloy ordering transitions for increased simulation cell sizes. We investigate the behavior of different models, with respect to their training capability as a function of the data set size while avoiding the danger of overfitting the model. An important aspect of our workflow is the periodic retraining with newly generated first principles data based on the progressive exploration of the system's phase space by the MC simulation. We will present results of these different approaches for training surrogate models for magnetic alloy materials.