Case Study: Eliminating Parallax Artefacts in X-ray Powder Diffraction Computed Tomography with a Self-Supervised Neural Network

Companies involved:

Finden Ltd, University College London, DESY, STFC

Challenge:

Parallax artefacts in X-ray powder diffraction computed tomography (XRD-CT) present significant challenges for accurately capturing physicochemical data from large samples. These artefacts, seen as peak shifting, broadening, and splitting in diffraction patterns, can lead to incorrect measurements of properties such as lattice parameters and crystallite sizes. Previously we developed a direct least-squares reconstruction (DLSR) algorithm to address this issue. However, that method required detailed chemical knowledge of the sample and was computationally intensive, making it unsuitable for large datasets increasingly common in modern research, particularly those acquired at synchrotron facilities.

Sample:

Commercial cylindrical Li-ion NMC532 battery

Solution:

Schematic representation of the parallax artefact and the improvement in image recon-struction after using ParallaxNet.

Figure 1: Schematic representation of the parallax artefact and the improvement in image recon-struction after using ParallaxNet.

In response to these challenges, we developed a new approach called ParallaxNet. This method leverages a 3D self-supervised neural network combined with a forward projector that takes experimental geometry and sample thickness into account. ParallaxNet eliminates the need for prior knowledge about the sample’s chemical composition, offering a more flexible and scalable solution compared to previous techniques. The SD2Vol neural network architecture shown in Figure 2 provides a framework for reconstructing parallax-free XRD-CT images without extensive data pre-processing. Tested on both simulated and experimental data, including a custom phantom sample and a commercial lithium-ion battery, ParallaxNet demonstrated its ability to accurately reconstruct diffraction images using a reduced scan range of 0°–180°.

Figure 2: a) The self-supervised ParallaxNet flow chart. The generator outputs a volume where the third dimension corresponds to the scattering angle. A forward operator is applied to convert these XRD-CT images into sinograms containing parallax artefacts. A loss function is then used to com-pare the differences between the generated sinograms and the experimental sinograms. b) The Single Digit to Volume (SD2Vol) generator architecture with a single constant as input. CONV rep-resents 3-D convolutional layers, and FC represents fully connected layers.

Figure 2: a) The self-supervised ParallaxNet flow chart. The generator outputs a volume where the third dimension corresponds to the scattering angle. A forward operator is applied to convert these XRD-CT images into sinograms containing parallax artefacts. A loss function is then used to com-pare the differences between the generated sinograms and the experimental sinograms. b) The Single Digit to Volume (SD2Vol) generator architecture with a single constant as input. CONV rep-resents 3-D convolutional layers, and FC represents fully connected layers.

Benefits:

The key benefits of the ParallaxNet approach include:

  • Elimination of Parallax Artefacts: ParallaxNet successfully removes artefacts such as peak shifting and broadening, ensuring accurate measurements of lattice parameters and crystallite sizes.
  • No Chemical Pre-Knowledge Required: Unlike DLSR, which necessitates a detailed physical model of the sample, ParallaxNet can reconstruct images without prior knowledge of the sample’s chemistry, reducing the complexity of the reconstruction process.
  • Scalability and Efficiency: ParallaxNet can handle larger datasets that traditional methods struggle with, making it suitable for high-resolution imaging applications. Additionally, by allowing for the use of a 0°–180° scan range, it cuts down the time required for data acquisition compared to conventional 0°–360° scans.
  • Simplified Workflow: The method operates directly on raw sinogram data without the need for extensive pre-processing, such as masking or background subtraction, streamlining the analysis process.

ParallaxNet’s versatility makes it applicable across a variety of fields, including material science, catalysis, and battery technology, where accurate tomographic reconstructions are critical for understanding complex systems.

Figure 3: a) The mean image of the Cu phase of the Li-ion battery dataset with a marked region of interest. b) Selected peak from average diffraction pattern from the region of interest. c) The lattice parameter a maps obtained by the Rietveld method for a Li-ion battery dataset. d) The distribution of lattice parameters for the maps is shown in (a).

Figure 3: a) The mean image of the Cu phase of the Li-ion battery dataset with a marked region of interest. b) Selected peak from average diffraction pattern from the region of interest. c) The lattice parameter a maps obtained by the Rietveld method for a Li-ion battery dataset. d) The distribution of lattice parameters for the maps is shown in (a).

Further reading:

Obtaining parallax-free X-ray powder diffraction computed tomography data with a self-supervised neural network. H. Dong, S.D.M. Jacques, K.T. Butler, O. Gutowski, A.-C. Dippel, M. von Zimmerman, A.M. Beale, A. Vamvakeros, npj Computational Materias, 10, 201, 2024. DOI: https://doi.org/10.1038/s41524-024-01389-1

 

Read more about our team and further publications.