Chemically disordered materials are at the heart of nuclear reactors with mixed uranium-plutonium oxides (MOX) being used as nuclear fuels and high-entropy alloys (HEA) having radiation-resistant properties that make them interesting candidates for structural materials. Predicting material properties using atomic scale simulation technics requires exhaustive sampling of the space of atomic configurations. This is particularly complex for disordered compounds, due to the multiplicity of possible atomic arrangements and the cost of atomistic simulations.

This work introduces a tool to explore with efficiency the configuration space of chemically disordered materials called Partition function Unsupervised Learning Sampling and Evaluation for disordered compounds (PULSE). Developed in our laboratory, it has proven its effectiveness in estimating local properties of mixed uranium-plutonium oxides [1]. In this work, we present its extensions to sample and estimate global properties such as thermal properties.

The PULSE method is based on a generative machine learning approach using a modified variational autoencoder model. The model iteratively generates atomic configurations, whose energy and forces are computed with conventional atomic-scale methods. The outcomes are looped back into the model to provide an estimate of the partition function, until convergence. Previous work showed that it worked for estimating local properties, i.e. measurable around a given atom in the simulation box such as defect formation energies[1]. We now extend it to calculations over the entire space. More specifically, it already we now apply it to the heat capacity of this MOX compounds.

We show that PULSE significantly reduces the computing time necessary to estimate atomic-scale properties of disordered compounds compared to other traditional methods. This study will open the way to a broader application of the PULSE method to thermal properties in MOX and HEA alloys.