GALAXI

GALAXI identifies the phases present in a powder X-ray diffraction pattern. Rather than a single multi-class model, it trains an independent classifier for each phase; the classifiers narrow a pattern to a few candidates, and Rietveld refinement settles which combination explains it. Because the classifiers are independent, a new phase is added by training one more of them, without retraining the rest.

The pretrained example catalog released upstream covers 365 phases drawn from battery and oxide chemistry rather than rock-forming minerals, so as shipped it does not identify arbitrary minerals. The lab holds GALAXI for training classifiers for the minerals it works with. A larger library of 64,594 structures sits behind upstream’s web service at galaxi-xrd.com; it is not distributed, and the Fossett Lab neither holds nor runs it.

autoXRD does the same job with a single model. The two have not been compared on the same patterns.

The code and the pretrained weights are both MIT-licensed.