Probabilistic Framework for Crystal Structure Denoising and Phase Classification

Researchers have developed a unified probabilistic framework designed to analyze noisy atomic configurations in atomistic simulations. Addressing the challenge of extracting phase labels and continuous order parameters from large volumes of structural data, this new model integrates thermal-noise removal, phase classification, and order parameter construction into a single differentiable scalar system. The framework predicts per-atom, per-prototype logits, aggregating them into a log-probability landscape where gradients define a conservative denoising field. Trained on AFLOW-mapped crystalline structures from the Materials Project with synthetic perturbations, the model demonstrates robust extrapolation capabilities. It successfully handles stronger noise, finite-temperature disorder, point defects, water-ice coexistence, binary polymorphs, and shock-compressed titanium. By recovering prototype identity after denoising and tracking smooth transformations like Bain and Burgers paths, the tool exposes low-confidence regions near defects and phase boundaries. This integrated approach offers an extensible solution for analyzing complex atomistic simulations, overcoming limitations of existing tools that often specialize in limited prototypes or separate processing steps.