Deep‐Learning‐Enhanced Electron Microscopy for Earth Material Characterization

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 2
• Main Authors: Melick, Hans, Taylor, Richard, Plümper, Oliver
• Format: Journal Article
• Language: English
• Published: Wiley 01.06.2025
• Subjects: deep learning; generative adversarial network (GAN); image registration; microstructure; scanning electron microscope (SEM); super resolution
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000549

Rocks, as Earth materials, contain intricate microstructures that reveal their geological history. These microstructures include grain boundaries, preferred orientation, twinning and porosity, holding critical significance in the realm of the energy transition. As they influence the physical strength, chemical reactivity, and transport and storage properties of rocks, they also directly impact subsurface reservoirs used for geothermal energy, nuclear waste disposal, and hydrogen and carbon dioxide storage. Understanding microstructures and their distribution is therefore essential for ensuring the stability and effectiveness of these subsurface activities. Achieving statistical representativeness often requires the imaging of a substantial quantity of samples at high magnification. To address this challenge, this research introduces a novel image enhancement process for scanning electron microscopy data sets, demonstrating significant resolution improvement through Deep‐Learning‐Enhanced Electron Microscopy (DLE‐EM). This workflow involves capturing high‐resolution (HR) regions within a low‐resolution (LR) area, and registering them with subpixel accuracy. First, the HR region's location is determined using a Fast Fourier Transform algorithm, followed by iterative refinement via a deformation matrix optimized with Newton's method to minimize image differences. The paired HR and LR images are then used to train a Generative Adversarial Network, where a generator and discriminator jointly train through an adversarial process to produce HR images from LR inputs. This approach accelerates imaging processes, up to a factor of 16, with minimal impact on quality and offers possibilities for real‐time super‐resolution imaging of unknown microstructures, promising to advance geoscience and material science. Plain Language Summary Rocks contain small‐scale structures, known as microstructures, that contain information about their history and physical properties. These microstructures, such as cracks, grain shapes and orientations, and pores, play a crucial role in how rocks behave. This is because microstructures control the strength of rocks, their ability to let fluids pass through, and their chemical reactivity. This makes microstructures important for activities such as geothermal energy production, hydrogen or carbon dioxide storage, and safely storing nuclear waste. To better understand the way microstructures are distributed in subsurface formations, scientists need to examine large numbers of rock samples in great detail, which can be time‐consuming and complex. This study presents a new method, called Deep‐Learning‐Enhanced Electron Microscopy (DLE‐EM), that relies on a novel image registration procedure to align high‐ and low‐resolution data which is then processed in a deep learning framework to enhance images from scanning electron microscopes. This allows microstructures to be studied up to 16 times faster without sacrificing image quality. Key Points This research introduces a novel deep‐learning‐based workflow to significantly accelerate scanning electron imaging of Earth materials Our integrated image registration and deep‐learning workflow achieves a 6‐ to 16‐fold increase in sample throughput in real‐world scenarios This integrated approach enables the capture of detailed microstructures across wide spatial areas, advancing the study of Earth materials