AlphaDiffract is a deep-learning model designed to analyze powder X-ray diffraction, a technique in which a material’s structure is inferred from the pattern produced by X-rays scattered by a powdered sample. It makes simultaneous predictions about the crystal’s broad symmetry, its more specific space group and the six lattice parameters that describe its repeating unit.
AI reads powder X-rays to map the structure of crystals
AlphaDiffract predicts crystal symmetry and all six measurements that describe a crystal’s repeating unit from diffraction patterns.

What the AI predicted
The model was trained on more than 31 million simulated diffraction patterns generated from 312,267 curated crystal structures in the ICSD and Materials Project databases. It uses a one-dimensional convolutional neural network with separate prediction components for crystal system, space group and lattice parameters. When tested on the predominantly experimental RRUFF benchmark, AlphaDiffract achieved 81.7% accuracy for crystal-system classification and 66.2% accuracy for space-group classification, while also predicting all six lattice parameters. Performance varied by crystal system: cubic materials were identified most accurately, while triclinic and monoclinic materials were more difficult.
Why one model matters
Determining a crystal’s lattice is often an early step toward solving its full structure. Existing approaches can be iterative and may require expert input, particularly when diffraction peaks are broadened, overlap or are affected by impurities. AlphaDiffract offers one model that makes the main lattice predictions together from a powder diffraction pattern, which could support faster analysis of materials datasets and experiments.
Evidence and caveats
This is a machine-learning study trained largely on simulated diffraction data and evaluated for transfer to experimental measurements using the RRUFF benchmark. The reported experimental-benchmark results show substantial, but incomplete, accuracy. The training data are unevenly distributed across crystal classes and space groups, with common groups overrepresented and rare groups less represented. Accuracy also varies considerably by crystal system, with lower-symmetry systems generally more challenging. The excerpts do not establish that the model can solve complete crystal structures or that it performs equally well across all experimental conditions.
// Source
npj Computational Materials · 2026 · DOI: 10.1038/s41524-026-02388-0
Authors: Nina Andrejevic, Ming Du, Hemant Sharma, James P. Horwath, Aileen Luo, Xiangyu Yin, Michael H. Prince, Brian H. Toby, Mathew J. Cherukara
Institutions: Argonne National Laboratory, Advanced Photon Source


