Answers are grounded in this paper and cite their sources.
The development of high-throughput x-ray diffraction (XRD) techniques presents significant challenges to traditional data analysis frameworks. The present work demonstrates that deep convolutional neural networks (CNNs) can reliably quantify key microstructural parameters of polycrystalline materials, specifically crystallite size and heterogeneous strain (microstrain) extracted from powder XRD patterns, within milliseconds. A CNN was trained using synthetic data generated with a dedicated code based on the GSAS-II software. We employ the example of monoclinic ZrO 2 to critically examine the theoretical performances of the CNN. The accuracy of the predictions was systematically investigated within a crystallite size range of 5–1000 nm and a microstrain range of 0.05%–2%. In this context, the role of the resolution of the diffractometer and the XRD peak profile shape is discussed in details. For instance, on a low-resolution laboratory diffractometer it is possible to determine the crystallite size within a 5–100 nm range with 99.7% accuracy in the absence of microstrain, and 97.3% accuracy when microstrains are present. Higher accuracies are obtained with a high-resolution diffractometer and over an extended range of size and microstrain. The potential of the CNN is demonstrated by the analysis of the crystallite size and microstrains in MgAl 2 O 4 formed by a solid-state reaction between MgO and Al 2 O 3 using in situ XRD at 1200 °C performed at the ESRF. The CNN can determine the crystallite size and microstrain in 0.004 s per diagram with values that deviate by ∼7% and ∼8% from a conventional Rietveld refinement, respectively, which is contained within the statistical uncertainties of these parameters. We further demonstrate that the same training dataset can be used for other regression problems, for instance phase fraction quantification, with no additional computational cost and only minor modifications to the network architecture. This work paves the way for real-time data analysis at synchrotron facilities.
No takes yet. Share an insight, caveat, or question.
Souesme et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: