Intelligent Detection of SEM Mineralogy Using Dynamic Segmentation Algorithm in Geothermal Sedimentary Reservoir: Case Study with Quantification of Quartz Overgrowth

Abstract

Abstract Quantifying quartz overgrowth in sedimentary geothermal reservoirs can provide vital information about reservoir quality and drilling success rates. The traditional process, which involves manual inspection of numerous Scanning Electron Microscope (SEM) images, is tedious and time-consuming. This paper introduces an automated approach using computer vision and random forest algorithms to streamline the process, providing a more efficient method for noise reduction, multi-level thresholding, machine learning (ML) model training, and application to SEM images for quartz overgrowth detection. Our approach employs a dynamic segmentation algorithm that integrates noise suppression, multi-level auto-thresholding, and dynamic overlaying. This enables automatic mineral identification from lower-quality images, accommodating varying brightness and contrast levels. Additionally, the algorithm effectively handles overlay shifting in cathodoluminescence (CL) and backscattered electron (BSE) images. We use a Random Forest technique to train the algorithm using comprehensive ground truth data and CL image features like Gabor, Canny Edge, and Roberts Edge. The resultant ML model helps refine the image segmentation predictions, acting as a more precise mineralogy predictor. Our model training achieved an encouraging 75% accuracy score, demonstrating its effectiveness in distinguishing between quartz grains and overgrowth, as well as identifying porosity and other minerals. However, due to the vast data size, further model improvement necessitates additional training data and usage of a High Performance Computing (HPC) cluster. After training, the model showed enhanced detection capabilities, particularly a reduction in false porosity detections associated with internal cracks. Future improvements could involve applying morphology detection principles and allowing user input for parameters like overlay shifting. The program was designed for easy use by individuals with minimal coding experience, and further development into a web app is suggested for increased accessibility. In summary, this method significantly enhances the efficiency and accuracy of mineralogy analysis using SEM images, providing a quick and accurate method for geoscientists to determine mineral composition.