Tackling real-world crystal structure prediction from powder X-ray diffraction data

Area of Science:

• Materials Chemistry
• Crystallography
• Computational Materials Science

Background:

• Crystal structure prediction (CSP) from powder diffraction data is a significant challenge in materials chemistry.
• Existing machine learning (ML) models for CSP are often trained on simulated data, limiting their reliability in real-world experimental settings.
• The deCIFer model, an autoregressive transformer, was developed for PXRD-conditioned generative CSP.

Purpose of the Study:

• To assess the real-world performance of the deCIFer ML model for powder X-ray diffraction (PXRD)-conditioned crystal structure prediction.
• To quantify the model's robustness against common experimental artefacts in PXRD data.
• To introduce metrics for evaluating accuracy and predictive uncertainty in ML-based CSP.

Main Methods:

• The deCIFer model, an autoregressive transformer, was used, conditioning structure generation on encoded PXRD data.
• Controlled robustness tests were performed using simulated artefacts: noise, background, peak asymmetry, and Scherrer broadening.
• Experimental PXRD data for known structures (Si, CeO2) and challenging cases (Fe2O3, nanocrystalline CeO2) were used for validation.

Main Results:

• The deCIFer model demonstrated smooth adaptation to signal distortions and outperformed unconditioned baselines when PXRD features were informative.
• The model effectively expressed predictive uncertainty as PXRD patterns became underdetermined.
• Experimental tests successfully recovered known structures and highlighted limitations for lower-symmetry and nanocrystalline materials.

Conclusions:

• ML-based CSP is fundamentally limited by the information content of PXRD data.
• ML models like deCIFer can accelerate expert workflows by generating plausible candidates and quantifying uncertainty.
• These models serve as valuable human-in-the-loop tools for real-world structure determination.