TorchDisorder: A Differentiable Framework for Generating Physically Realistic Disorder Structures from Experimental...

Determining the atomic structure of amorphous materials remains a fundamental challenge in condensed-matter physics and materials science. Unlike crystalline solids, disordered systems lack long-range periodicity, which makes conventional diffraction analysis insufficient for resolving the three-dimensional atomic arrangements. Existing reverse Monte Carlo (RMC) approaches rely on stochastic sampling, limiting both computational efficiency and the ability to enforce chemically realistic local environments. Here we present TorchDisorder, a PyTorch-based framework that replaces stochastic moves with gradient-based optimization via automatic differentiation, built on three tightly integrated components: GPU-accelerated neighbor list construction via torch-sim, augmented Lagrangian constrained optimization via the Cooper library, and a differentiable structure factor engine that propagates gradients through the full Faber-Ziman weighted Fourier transform. Coordination constraints for tetrahedral, octahedral, and other geometries are specified via JSON configuration files generated automatically from crystalline precursors and require no manual parameter tuning. We apply TorchDisorder to three glass systems relevant to energy technology, namely, silica (SiO₂), germania (GeO₂), and lithium thiophosphate (Li₂S-P₂S₅) solid electrolytes, and obtain structural models in quantitative agreement with experimental scattering data (R² ≥ 0.955) within 5000 gradient steps using a single diffraction data set per system, outperforming stochastic RMC in both convergence speed and constraint satisfaction.