Speaker
Description
Powder diffraction is a notoriously ill-posed inverse problem: the observed one-dimensional pattern entangles lattice parameters, space group symmetry, atomic arrangement, microstructural effects, and instrumental contributions, which classical pipelines address sequentially with limited reproducibility. We present CrystalAI, a machine learning framework for end-to-end crystal structure solution from powder diffraction patterns. At its core is a dedicated disentanglement block that learns to separate the physical contributions to the diffraction signal, coupled with crystallographic constraints embedded into the architecture and training. A central design principle is joint training on simulated and real experimental patterns: simulation alone leaves a persistent sim-to-real gap, while experimental data alone is too scarce. We present models for crystal system and space group prediction, together with a generative component proposing atomic positions in the unit cell. To make these models accessible to the crystallographic community, we are integrating them into AIXtal, a Rust-based web application for powder diffraction analysis and refinement.
This integration makes CrystalAI directly relevant to the FAIR data agenda advanced by DAPHNE4NFDI. Such models depend on curated, openly accessible diffraction data and produce reusable artifacts — learned priors, trained models, structured outputs — that themselves require FAIR stewardship.
