Skip to main content
Search

Ming Du

1 related articles

Subscribe to updates
  1. AlphaDiffract reads a powder pattern once: 81.7% crystal-system accuracy on experimental RRUFF data, with all six lattice parameters at the same time

    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.

You’re all caught up