Deep Learning-Based Segmentation for Quartz Overgrowth Quantification in Geothermal Sandstone Reservoirs from Scanning Electron Microscope (SEM) Images

Abstract

Abstract This paper presents an update on a dynamic segmentation algorithm for detecting quartz overgrowths in geothermal reservoirs using Scanning Electron Microscope (SEM) images. Previously, the Random Forest algorithm had been employed in the automated workflow for quartz overgrowth detection from SEM images. A 75% accuracy score was achieved from the model training, indicating a promising start. This model was found to differentiate successfully between detrital quartz grains and their diagenetic quartz overgrowths; it was also demonstrated that it could identify porosity and other minerals. A continuation of the algorithm development in the automated workflow is explored in this paper. Deep learning methods using U-net architecture is investigated to find the most fitting algorithm for detecting quartz overgrowth. The previously utilized texture-based feature extraction techniques are still incorporated. Normalization and dynamic overlaying algorithms are applied to address variations in image brightness and contrast and align BSE and CL images accurately, ensuring reliable segmentation. The segmentation process involves the coordination of BSE and CL images, utilizing their respective strengths, and overlaying them to achieve comprehensive results. This is followed by a two-fold model-building approach using separate segmentation models for BSE and CL images, which are then combined to distinguish between pore space, quartz grain, and quartz overgrowth. The evaluation of the U-Net model's performance involves analysis of training and validation accuracy, loss, and intersection over union (IoU) over 50 epochs. Results demonstrate the model's capability to generalize and learn effectively, with the segmentation process showing proficiency in differentiating between the target mineralogy features. However, variability in performance across different datasets suggests the need for further model optimization. In conclusion, the integration of U-Net into SEM image analysis for mineralogy detection represents a significant technological advance in geoscience, offering a more efficient, precise, and automated approach to understanding and exploiting geothermal energy resources. The findings also highlight opportunities for future research, such as exploring a variety of deep learning models, fine-tuning through transfer learning, and developing user-friendly tools for rapid mineralogy segmentation.