Breaking Boundaries: Deep-Learning-Enhanced Electron Microscopy for Accelerated Super-Resolution Imaging in Solid Earth Research

In the realm of modern solid Earth research, a profound understanding of rocks' intricate microstructures is essential for unraveling geological history and addressing critical challenges in the energy transition. These microstructures—grain boundaries, preferred orientation, twinning, and porosity—play a pivotal role, influencing the physical strength, chemical reactivity, and fluid flow properties of rocks. Their direct impact on subsurface reservoirs used in geothermal energy, nuclear waste disposal, and hydrogen/carbon dioxide storage underscores the importance of comprehending their distribution for the stability and efficacy of subsurface activities. However, addressing the need for statistical representativeness requires imaging numerous samples at high magnification. In response, our research introduces an innovative image enhancement process for scanning electron microscopy datasets, showcasing a substantial potential for resolution improvement through Deep-Learning-Enhanced Electron Microscopy (DLE-EM). Our proposed workflow involves capturing one or more high-resolution (HR) regions within a low-resolution (LR) area. Precise image registration is achieved in two steps: first, determining the HR region's location within the LR region using a Fast Fourier Transform algorithm (Lewis, 2005), and second, refining image registration through iterative calculation of a deformation matrix. This matrix, utilizing Newton's optimization method, aims to minimize differences between both images (Tudisco et al., 2017). Subsequently, paired HR and LR images undergo processing in a Generative Adversarial Network (GAN), comprising a generator and a discriminator. This GAN learns to generate HR images from LR counterparts through joint training in an adversarial process. We benchmark our workflow using four distinct rock types and demonstrate that this approach accelerates imaging processes up to a factor of 16 with minimal impact on quality, offering possibilities for real-time super-resolution imaging of unknown microstructures. Additionally, we show that a model trained on a specific geological material is able to generalize its learned features to new domains, reducing the need for extensive training data. [1] Lewis, J. P. "Fast normalized cross-correlation, Industrial Light and Magic." unpublished (2005). [2] Tudisco, Erika, et al. "An extension of digital volume correlation for multimodality image registration." Measurement Science and Technology 28.9 (2017): 0954