Physics-constrained machine learning for powder x-ray diffraction

Powder x-ray diffraction is the workhorse of crystalline materials characterization, yet its interpretation still relies heavily on expert judgment. Machine learning offers a route to automate this process, with progress over the past decade revealing a clear trend toward increasingly reliable models as more physical knowledge is incorporated. Early neural networks treated diffraction patterns as generic one-dimensional signals and often failed to generalize from simulated to experimental data. Performance improved as diffraction physics was incorporated first into the data and then into the models used to interpret them. In this Perspective, we trace this progression and argue that the next step is to constrain inference using computational priors. Machine-learned interatomic potentials now make it inexpensive to predict quantities such as phase stability, thermal expansion, and defect behavior, which can constrain which phases are likely to form and how their diffraction patterns evolve with changing experimental conditions. Incorporating these priors directly into the inference objective could improve the reliability of both ex situ analysis and in situ or operando measurements, and we discuss the opportunities and challenges involved.