Characterizing the atomic structure of functional materials is essential for advancing technologies in energy storage, catalysis, and electronics. Powder X-ray diffraction (PXRD) remains a central tool for this task, providing average periodic structure even for nanomaterials where finite size and surface effects are present. Yet identifying a suitable crystallographic model directly from PXRD is difficult, as instrumental resolution limits and sample characteristics broaden or distort peaks, creating overlap and ambiguity. In parallel, generative and machine learning approaches to crystal structure prediction have progressed rapidly, but most operate on high-level descriptors such as composition or symmetry and do not incorporate diffraction data directly^1,2,3.This motivates a central question: how does PXRD conditioning shape the accuracy and uncertainty of generative crystal structure predictions?

To investigate this, we introduce deCIFer⁴, a PXRD-conditioned autoregressive transformer that generates full CIF structures from an encoded diffraction profile, with optional inputs such as composition or space group. deCIFer is trained on a large set of inorganic structures paired with simulated PXRD patterns and therefore provides a representative case study for PXRD-guided generative modelling.

We assess robustness through a unified evaluation combining synthetic perturbation experiments and real experimental PXRD. Synthetic tests apply physically motivated distortions to the input pattern, including unit-cell scaling, peak shifts, peak asymmetry, Scherrer broadening, additive noise and background variation. Across these conditions, we observe that PXRD conditioning improves structure prediction when diffraction features are sufficiently informative, producing tight sets of plausible structural candidates. As distortions increase, deCIFer transitions to broader structural distributions that reflect the diminishing information content of the PXRD, and when the signal becomes uninformative, the predictions revert toward statistically favored structures. Lastly, we evaluate the model on experimental PXRD from well-characterized bulk- and nanomaterial samples. In these cases, deCIFer provides reasonable structural candidates that capture the main diffraction features that can serve as starting models for downstream analysis. These results highlight that PXRD conditioning can improve the robustness of crystal structure prediction and support practical workflows in materials characterization.