Super-resolving 3D nanostructures using artificially generated image data and spatial transport simulations

Abstract An approach for deploying stochastic 3D models to generate microstructural 3D image data for training super-resolution networks is investigated for three different scaling factors α∈{2,4,8}. The presented approach addresses the issue of scarcity in training data by training the networks only on artificial image data, generated by means of a stochastic 3D model that produces digital twins of the nanoporous inner structure of active particles in battery cathodes. In addition, the performance of super-resolution networks is investigated when complementing the input data, i.e., low-resolved microstructural 3D image data, with spatially resolved transport simulations. The performance of the trained networks is evaluated based on real tomographic image data, and quantified with respect to various geometric descriptors and effective transport properties. It turned out that the integration of transport simulations into the training of super-resolution networks showed an increase in performance for the scaling factorsα∈{2,4}, but a decrease in performance for α=8. However, training the networks on artificial image data was effective in all cases.