Effective Property Uncertainty of Graphite Electrodes from Computed Tomography Using Bayesian Convolutional Neural Networks

In recent years there have been great advances in using 3D tomography in battery research, providing the ability to see as-manufactured electrodes at the mesoscale. These structures are subsequently used to find effective properties or perform detailed electrochemical simulations. A key step of this analysis is image segmentation, which differentiates particle and void phases. Image segmentation, however, is painstaking and fraught with subjectiveness. Removing this subjectiveness is paramount in having credible direct numerical simulations to bridge the gap of the computational and experimental regimes. To improve the image segmentation process and quantify its uncertainty, a Bayesian convolutional neural network (BCNN) is implemented to segment greyscale tomograms of graphite electrodes. The use of a BCNN allows for native and explainable uncertainty estimates. Two manually segmented tomograms were used to train the BCNN, which was in turn used to segment an additional four tomograms. This process was effective in creating high quality binary stacks, which in some cases presented a subjectively superior segmentation compared to the manual process. Additionally, BCNN-generated uncertainty estimates provide a range of possible image segmentations and structures. These nominal and uncertainty ranged structures are propagated through simulations of effective electrical conductivity and pore phase tortuosity, providing both nominal property predictions and geometric uncertainty estimates of those predictions. This work demonstrates a new method for quantifying geometric uncertainty from as-manufactured electrode images and the impact of that uncertainty regarding battery-relevant physical phenomena.