AI-driven data-efficient estimation of partition functions in disordered materials

This work presents PULSE (Partition function Unsupervised Learning Sampling and Evaluation), a novel generative approach to efficiently estimate the partition function of disordered compounds, without requiring pre-existing training datasets and at a fraction of the computational cost compared to traditional methods. By enabling targeted calculations of atomic-scale properties, PULSE overcomes key limitations of existing sampling techniques, which either require prohibitive computational effort (Monte Carlo) or result in a partial exploration of the configuration space (Special Quasirandom Structures). The approach, based on an inverse variational autoencoder architecture, can also generate representative configuration sets tailored to specific properties, offering a powerful tool for constructing optimized datasets for training interatomic potentials. We demonstrate the capabilities of our method by computing point-defect formation energies and concentrations in uranium-plutonium mixed oxides. Furthermore, we show that the generated configurations provide physical insight into the role of local environments on defect behavior. Beyond this specific application, PULSE is broadly applicable to studying disorder-driven properties in complex materials, including mixed oxides and high-entropy alloys.