AlphaDiffract reads a powder pattern once: 81.7% crystal-system accuracy on experimental RRUFF data, with all six lattice parameters at the same time
Synopsis
Recovering a three-dimensional unit cell from a one-dimensional powder X-ray diffraction pattern has meant iterative indexing that depends on expert input, and most deep-learning attempts classify only crystal system and space group or train separate models per crystal system. AlphaDiffract pairs a one-dimensional ConvNeXt backbone with three parallel prediction heads to output crystal system, space group and all six lattice parameters from a single pattern, trained on more than 31 million simulated patterns generated from 312,267 structures. On the largely experimental RRUFF benchmark, a ten-model ensemble reaches 81.7% crystal-system accuracy and 66.2% space-group accuracy, with a mean absolute error of 2.11 angstroms on lattice lengths.
Interpretation
AlphaDiffract puts crystal system, space group and six lattice parameters into one inference pass whose only input is a powder diffraction pattern, with no chemical formula and no model trained separately per crystal system. Most prior work only classifies crystal system and space group; approaches that predict lattice parameters either train separate models for each Bravais lattice or crystal system, require prior symmetry knowledge, depend on chemical composition as input and need subsequent refinement for experimental data. On the RRUFF benchmark the ten-model ensemble reaches 81.74% crystal-system accuracy (95% confidence interval 78.88–84.60) and 66.21% space-group accuracy (62.81–69.62), against 26.98% and 3.27% for a naive baseline that always predicts the most abundant training class.
The model first extracts shared features from the pattern, then three prediction heads that do not communicate separately output crystal system, space group and lattice parameters, so lattice-parameter prediction is not hard-conditioned on a possibly wrong crystal-system label. The backbone is a one-dimensional ConvNeXt, a convolutional network that folds in transformer design elements such as depthwise separable convolutions, an inverted bottleneck and large kernels; average pooling after each block widens the receptive field and yields a 560-dimensional feature vector. Training patterns were simulated with GSAS-II from 312,267 ICSD and Materials Project structures, 100 per structure, with randomized microstrain, crystallite size and instrumental broadening plus Poisson and Gaussian noise; RRUFF patterns were mapped by Bragg's law into the 2θ range corresponding to 20 keV before evaluation.
Perspective
The demonstrated setting is single-phase powder diffraction patterns: training and synthetic evaluation use simulated patterns at 20 keV over 2θ 5°–20° with 8192 points, and RRUFF measured patterns are mapped by Bragg's law into the same d-spacing range before evaluation; classification uses 734 structures, while lattice-parameter regression uses only the 240 of them with complete atomic coordinates. The authors state the model does not replace phase-identification software such as JADE or EVA, which match patterns against databases of known phases, but targets unknown or novel materials with no database entry.
The authors note that roughly 20% mean absolute percentage error on lattice lengths is not yet sufficient to initialize Pawley or Le Bail whole-pattern fits, and their attempts to do so did not converge; the model does not explicitly handle preferred orientation (texture) and does not model planar defects such as twinning and stacking faults, and space-group accuracy on measured patterns is 15–17% lower than on simulated patterns from the same structures. On the subset excluding structures equivalent to training data, RRUFF space-group accuracy falls to 49.82%, which the authors stress uses a stricter deduplication criterion than prior studies and should not be substituted into the original comparison tables.
