Automated prediction of lattice parameters from X-ray powder diffraction patterns

Area of Science:

• Materials Science
• Crystallography
• Computational Chemistry

Background:

• Accurate determination of unit-cell lattice parameters is crucial for powder X-ray diffraction (PXRD) data analysis.
• Current methods often require extensive human intervention, creating a bottleneck for automated analysis.

Purpose of the Study:

• To develop machine learning models for automated lattice parameter determination.
• To assess the impact of experimental non-idealities on model performance.
• To integrate machine learning with existing refinement schemes for improved automated unit-cell solution.

Main Methods:

• Development and training of one-dimensional convolutional neural networks (1D-CNNs) for lattice parameter estimation.
• Training on a large dataset of nearly one million crystal structures from ICSD and CSD databases.
• Systematic analysis of experimental non-idealities like impurity phases, noise, and peak broadening.

Main Results:

• Achieved a mean absolute percentage error of approximately 10% for lattice parameter estimation across crystal systems.
• Demonstrated a 100- to 1000-fold reduction in the lattice parameter search space volume.
• Identified impurity phases, baseline noise, and peak broadening as major challenges, while zero-offset and intensity modulations had minimal impact.

Conclusions:

• 1D-CNN models offer a robust approach to automated lattice parameter determination in PXRD.
• Data augmentation strategies can mitigate the effects of experimental non-idealities.
• Combining machine learning estimates with iterative refinement schemes enables accurate automated unit-cell solutions.