Machine Learning Method Decomposes Multi-Phase X-ray Diffraction Patterns From Single Observation

Powder X-ray diffraction is a standard technique used for structural characterization of crystalline compounds. Analysis becomes difficult when samples contain multiple phases, producing overlapping diffraction patterns that must be separated before further interpretation.

Traditional decomposition methods typically need either several mixture samples with varying compositions or advance knowledge of the expected phases. These requirements limit their use in complex systems or high-throughput experiments where such information is unavailable.

The method, named Phase Decomposition Diffusion Transformers or PhaseDifformer, treats the denoising process in diffusion models as a probabilistic regressor to recursively extract unknown constituent phases. The researchers tested the approach on both synthetic mixtures and real experimental measurements.

According to the study, the model achieved accurate phase decomposition in both settings.

The development fills a gap in automated X-ray diffraction workflows. Existing single-phase methods can convert clean diffraction patterns to crystal structures, but mixed-phase patterns have required manual intervention or restrictive assumptions. By enabling decomposition from one observation, the technique supports broader application in materials research where multi-phase systems are common.

The authors noted that the work advances toward fully automated end-to-end analysis of complex diffraction data. The research received support from multiple Japanese funding programs including the JST-Mirai Program and the JST Moonshot R&D Program. Additional grants came from JSPS KAKENHI and the MEXT Program for material research and development.