Fusion of complementary 2D and 3D mesostructural datasets using generative adversarial networks

TL;DR

This paper presents SuperRes, a GAN-based framework that fuses complementary 2D and 3D imaging to reconstruct high-resolution, representative 3D mesostructures. By extending SliceGAN to accept low-res 3D inputs and incorporating a voxel-wise loss and adaptive blurring, it performs simultaneous super-resolution, style transfer, and dimensionality expansion across isotropic and anisotropic materials. Four case studies demonstrate accurate recovery of key mesostructural metrics and TBP densities, including a real-world demonstration with no available ground-truth high-res volumes. The method enables high-fidelity mesoscale simulations with open-source data and code, offering a practical route to overcome imaging limitations in energy materials and beyond.

Abstract

Modelling the impact of a material's mesostructure on device level performance typically requires access to 3D image data containing all the relevant information to define the geometry of the simulation domain. This image data must include sufficient contrast between phases to distinguish each material, be of high enough resolution to capture the key details, but also have a large enough field-of-view to be representative of the material in general. It is rarely possible to obtain data with all of these properties from a single imaging technique. In this paper, we present a method for combining information from pairs of distinct but complementary imaging techniques in order to accurately reconstruct the desired multi-phase, high resolution, representative, 3D images. Specifically, we use deep convolutional generative adversarial networks to implement super-resolution, style transfer and dimensionality expansion. To demonstrate the widespread applicability of this tool, two pairs of datasets are used to validate the quality of the volumes generated by fusing the information from paired imaging techniques. Three key mesostructural metrics are calculated in each case to show the accuracy of this method. Having confidence in the accuracy of our method, we then demonstrate its power by applying to a real data pair from a lithium ion battery electrode, where the required 3D high resolution image data is not available anywhere in the literature. We believe this approach is superior to previously reported statistical material reconstruction methods both in terms of its fidelity and ease of use. Furthermore, much of the data required to train this algorithm already exists in the literature, waiting to be combined. As such, our open-access code could precipitate a step change by generating the hard to obtain high quality image volumes necessary to simulate behaviour at the mesoscale.

Figures (18)

• Figure 1: The SuperRes model inputs and output evaluated on the isotropic NMC cathode XCT dataset (Case study 1) usseglio2018resolving, with a scale factor of 4. The different phases are pore (black), active material (gray) and binder (white). The top row shows the inputs and output of the model. For training, the model requires a high-res multi-phase 2D image and a low-res binary 3D volume ($\times 4$ stretched to make comparison to the super-res version easier), and the output is a super-res volume of the low-res volume with the same multi-phase characteristics of the high-res 2D image. The bottom row shows a comparison between a random cross section in the same position of the low-res and super-res volumes, and as well as the middle region of the high-res slice to compare the nature of the fine details.
• Figure 2: Statistical comparison of key mesostructural metrics between the original high-res 3D volume (blue), the high-res 2D slice (black) and the reconstructed super-res 3D volume (orange). The left figure shows a comparison of the volume fraction of the different phases in the material. The middle figure shows a comparison of the interphase surface area, the fraction of the number of neighboring voxels of different phases by the total number of faces in the volume. The right figure shows a comparison of the transport efficiency for each phase calculated using a diffusivity simulationcooper2016taufactor. All metric calculations were obtained using the TauFactorcooper2016taufactor package. All figures show a violin plot distribution of the 3D high-res and the 3D super-res taken from 256 randomly sampled $64^3$ volumes, with normalised width for each distribution and corresponding horizontal lines for the means. The black pluses mark the metrics measured on the whole 2D image. Note that since transport efficiency is a volumetric measurement, only the values for the 2D high-res slice are not included.
• Figure 3: The results from exploring the effect of different scale factors. For each scale factor, the inputs for the model were a high-res 2D slice, whose center can be seen on the right, and a low-res volume which was blurred and down-sampled accordingly to the scale factor from the original high-res volume. Every column shows a different experiment for a different scale factor. The top row shows a cross section of the original volume, the middle row shows a cross section of a corresponding position in the low-res volume and the bottom row shows a cross section of a corresponding position in a generated super-res volume.
• Figure 4: The model results for an anisotropic material, here producing a super-res 3D battery separator material finegan2016characterising with a scale factor of 4. The different model inputs and outputs are the same as in Figure \ref{['fig:Fig1-NMCimages']}, with the additional input of high-res 2D slices from all three perpendicular directions. Since two perpendicular directions have the same properties (upper slice in the top left corner), for simplicity only two high-res 2D slices are shown out of the 6 facets used for training data.
• Figure 5: Statistical comparison of key mesostructural metrics between the original high-res 3D volume, the high-res 2D slices and the reconstructed super-res 3D volume for the anisotropic separator material in Figure \ref{['fig:separator_outline']}. The metrics description is the same as in Figure \ref{['fig:nmc_metrics']}. Since the material is anisotropic, transport efficiency measurements were taken along all different axis. The figure also shows the transport efficiency measurements of the low-res volume, and the added information gained by super-resolving.