Title:Fast and stable predictions of total energy of solid solution alloys

Abstract:We present a deep learning approach to produce highly accurate predictions of macroscopic physical properties of solid crystals. Since the distribution function of the total energy would enable a thorough understanding of the macroscopic properties for alloys and magnetic systems, surrogate deep learning models can replace first principle calculations to speed up each sample from the total energy distribution function. However, neural network models lose accuracy with respect to first principle calculations and this affects the reliability of the estimate. Here, we focus on reducing the uncertainty of multitasking neural network models to simultaneously predict physical properties of alloys. The uncertainty is reduced relying on the physical correlation of the magnetic moment with total energy and charge density. These physical quantities mutually serve as constraints during the training of the multitasking deep learning model, which leads to more reliable deep learning models because the physics is also learned correctly. We present numerical experiments for two types of binary alloys: copper-gold and iron-platinum. The dataset comprises information about total energy, charge density and magnetic moment (for magnetizable materials) computed via first principle codes for 32,000 configurations which differ per composition to properly span the entire state space. Results show that multitasking neural networks estimate the material properties for a specific state space hundreds of times faster than the first principle codes used as a reference. Moreover, the inclusion of the magnetic moment as a physical constraint for iron-platinum significantly reduces the uncertainty of the predictions by damping the fluctuations of the predictive performance with respect to different validation splittings of the dataset.