A recent study published in the Journal of Geophysical Research: Machine Learning and Computation presents a new image enhancement process for scanning electron microscopy (SEM) data sets, called Deep-Learning-Enhanced Electron Microscopy (DLE-EM). DLE-EM significantly improved resolution and accelerated imaging speeds by up to 16 times.
Image Credit: Gorodenkoff/Shutterstock.com
Background
Earth materials such as rocks contain complex microstructures—grain boundaries, preferred orientations, twinning, and porosity—that record geological history and influence subsurface reservoirs used for nuclear waste disposal, geothermal energy, and hydrogen and carbon storage. Accurately characterizing these microstructures is critical for assessing the effectiveness of subsurface engineering activities.
While optical, electron, and X-ray microscopy techniques are used to study rock microstructures, they face limitations in time and resolution, particularly when capturing statistically meaningful sample sizes during deformation studies. Deep learning offers a way to overcome these challenges.
This study proposes DLE-EM as a method to enhance SEM images and improve throughput for geological analysis.
Methods
The researchers developed a Python-based workflow to collect large data sets of rock microstructures. The approach involved acquiring one or more high-resolution (HR) regions within a low-resolution (LR) SEM image.
The first step was template matching, identifying the location of the HR region within the LR map using normalized cross-correlation and Fast Fourier Transform techniques for efficiency. After locating the HR region, a two-step registration process minimized image disparities by optimizing a deformation matrix for pixel-level alignment.
Once images were registered, a Generative Adversarial Network (GAN) was trained to construct HR outputs from LR inputs. Training involved 46–300 epochs using a mean squared error (MSE) loss function, with an additional mean absolute error (MAE) training applied for Berea sandstone models.
Berea sandstone was used as a case study and primary benchmark to gauge the effectiveness of the proposed DLE-EM. Meanwhile, the generalization proficiency of the trained benchmark model was evaluated using data sets of three other distinct rock samples: gabbro, serpentinite, and schist.
Results and Discussion
Training outcomes varied significantly between the two loss functions. Models trained with MSE converged faster, were more stable, and showed no visual artifacts. In contrast, models trained with MAE exhibited dual modes and higher noise, consistent with fluctuations in discriminator loss.
After 50 epochs (MSE) and 150 epochs (MAE), both models effectively captured key geological features such as inclusions, grain boundaries, veins, cracks, and textures. The MSE model reached optimal performance sooner, while the MAE model required extended training.
Imaging a standard geological thin section at 50 nm resolution with a 3 µs dwell time typically takes over 17 days. Using the DLE-EM workflow, where only 10% of the sample is scanned at high resolution and the rest upscaled, a sixfold increase in throughput was achieved, reducing scanning time to under 3 days.
In the Berea sandstone case, beam time was cut by nearly a factor of nine. When applying a pre-trained model to new samples, high-resolution scanning was not required for every dataset, streamlining the process and boosting throughput by up to 16 times.
Download your PDF copy now!
Conclusion
The DLE-EM workflow demonstrated the ability to produce high-resolution, large field-of-view SEM images, enabling detailed characterization of complex rock microstructures across broader spatial scales.
This method improves statistical representation of geological features, allowing better predictions of fluid flow, reservoir behavior, and mechanical properties. DLE-EM has clear applications in subsurface engineering fields such as carbon and hydrogen storage, geothermal energy, and studies of fluid-rock interaction, seismicity, and geological evolution.
Journal Reference
Melick, H. van, Taylor, R., & Plümper, O. (2025). Deep‐Learning‐Enhanced Electron Microscopy for Earth Material Characterization. Journal of Geophysical Research Machine Learning and Computation, 2(2). DOI: 10.1029/2024jh000549, https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2024JH000549
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.