Reconstruction of incomplete X-ray diffraction pole figures of oligocrystalline materials using deep learning

David Meier, Rishan Ragunathan, Sebastian Degener, Alexander Liehr, Malte Vollmer, Thomas Niendorf, Bernhard SickScientific 2023 · DOI

TL;DR

Researchers have developed a deep learning framework capable of reconstructing complete X-ray diffraction (XRD) pole figures from a single, incomplete measurement. By leveraging a high-speed GPU simulator and a custom "standardization" neural network, the team reduced experimental measurement time for oligocrystalline materials by approximately 74% while maintaining high reconstruction accuracy.

What Problem Does This Paper Address?

Determining the crystallographic orientation of grains within a material is vital for understanding its mechanical properties, especially in high-performance materials like shape memory alloys (SMAs). While X-ray diffraction (XRD) is a preferred non-destructive method, it is notoriously slow. To create a "pole figure"—a graphical representation of crystal orientation—researchers must rotate and tilt a specimen through hundreds of positions, often across multiple lattice planes ($hkl$).

This process is particularly challenging for oligocrystalline materials (those with only a few large grains). Because these materials produce sparse diffraction spots rather than continuous rings, standard mathematical reconstruction methods often fail or require excessive data. The authors sought to bridge this gap: can we use deep learning to "fill in the blanks" of a partially measured pole figure, using data from only one lattice plane instead of three?

Methods

The authors implemented a three-stage pipeline to solve the reconstruction problem:

  1. GPU-Based Simulation: To train deep learning models, one needs massive amounts of data. The team built a custom GPU-accelerated simulator that generates pole plots from crystal orientations (represented as quaternions). This allowed them to generate over 250 million synthetic training samples without the bottleneck of physical experiments.
  2. Pole Width Standardization: Experimental data is "messy"—factors like beam divergence or material strain cause diffraction spots (poles) to vary in width and shape. The authors used a U-Net-like autoencoder to standardize these poles to a uniform width. This pre-processing step makes the subsequent reconstruction task much more robust against experimental noise.
  3. Reconstruction Network: The core architecture consists of a convolutional encoder (ConvEncoder) to extract features, a transposed convolutional decoder (T-ConvDecoder) to upsample the data, and a multilayer perceptron (MLP-Decoder) to predict the intensities of the unmeasured regions.
  4. Uncertainty Estimation: They employed Monte Carlo dropout to provide a spatial map of where the model is uncertain about its predictions, giving researchers a "confidence score" for the reconstructed data.

Datasets

  • Synthetic Dataset: 250 million simulated pole plots for the 211 lattice plane, featuring 1 to 3 grains with random orientations and varying pole widths.
  • Real-World Benchmark: A physical specimen of a Fe–Mn–Al–Ni–Ti shape memory alloy. The specimen was measured up to a 76° tilt to provide a "ground truth" label, which was then truncated to 60° to test the model’s reconstruction capabilities.

Key Insights and Ideas

  • Decoupling Noise from Geometry: One of the most clever aspects of this paper is the "Standardization" step. By training a specific network just to fix pole widths, the "Reconstruction" network can focus entirely on the geometric relationship between crystal orientations, making the whole system more accurate.
  • The "Sweet Spot" of Tilt: The study found that measuring up to a 60° tilt provides a critical mass of information. Beyond 60°, the reconstruction error decreases only slightly, suggesting that researchers can stop measurements early with minimal loss of information.
  • Simulation as a Necessity: The authors highlight that without their high-speed simulator, this machine learning approach would be impossible, as collecting 250 million real XRD patterns would take lifetimes of lab work.

Key Results

  • Massive Time Savings: In a real-world test on a shape memory alloy, the method saved 504 minutes (~74%) of measurement time. This was achieved by reducing the required tilt range and, crucially, needing only one lattice plane measurement instead of the usual three.
  • High Robustness: The standardization model maintained its performance even when 20% Gaussian noise was added to the input data.
  • Efficiency: The entire inference pipeline (standardization + reconstruction) runs in under 90 ms on a standard laptop CPU, allowing for real-time feedback during experiments.
  • Qualitative Accuracy: As seen in the paper's figures, the network successfully restored missing poles even in regions the X-ray detector never "saw," including poles with low intensity.

Limitations and Future Work

  • Grain Count Scaling: The current model is optimized for 1–3 grains. While this covers many specialized alloys, it is not yet applicable to fine-grained polycrystalline materials.
  • The MSE Metric: The authors acknowledge that Mean Squared Error (MSE) isn't the perfect metric for spherical data because it doesn't account for the "wrap-around" nature of stereographic projections. They suggest exploring the Wasserstein metric (Earth Mover’s Distance) in future versions.
  • Active Experimentation: The authors plan to use the uncertainty maps to create an "active" system where the XRD machine automatically identifies and probes regions where the model is most confused.

Summary updated

Abstract

X-ray diffraction crystallography allows non-destructive examination of crystal structures. Furthermore, it has low requirements regarding surface preparation, especially compared to electron backscatter diffraction. However, up to now, X-ray diffraction has been highly time-consuming in standard laboratory conditions since intensities on multiple lattice planes have to be recorded by rotating and tilting. Furthermore, examining oligocrystalline materials is challenging due to the limited number of diffraction spots. Moreover, commonly used evaluation methods for crystallographic orientation analysis need multiple lattice planes for a reliable pole figure reconstruction. In this article, we propose a deep-learning-based method for oligocrystalline specimens, i.e., specimens with up to three grains of arbitrary crystal orientations. Our approach allows faster experimentation due to accurate reconstructions of pole figure regions, which we did not probe experimentally. In contrast to other methods, the pole figure is reconstructed based on only a single incomplete pole figure. To speed up the development of our proposed method and for usage in other machine learning algorithms, we introduce a GPU-based simulation for data generation. Furthermore, we present a pole widths standardization technique using a custom deep learning architecture that makes algorithms more robust against influences from the experiment setup and material.
BibTeX
@inproceedings{meier2023reconstruction,
  author = {David Meier and Rishan Ragunathan and Sebastian Degener and Alexander Liehr and Malte Vollmer and Thomas Niendorf and Bernhard Sick},
  title = {{Reconstruction of incomplete X-ray diffraction pole figures of oligocrystalline materials using deep learning}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-31580-1},
}
Figure