| Title | Dynamic Segmentation ML Algorithm for Inferring Geothermal Reservoir Quality in Sandstones using Scanning Electron Microscope (SEM) Images: Case Study with Quantification of Quartz Overgrowths |
|---|---|
| Authors | Sarah Sausan, Arkanu Andaru |
| Year | 2023 |
| Conference | World Geothermal Congress |
| Keywords | Mineral segmentation; backscattered electron; cathodoluminescence; scanning electron microscope; geothermal reservoirs. |
| Abstract | The abundance of diagenetic quartz overgrowths in sedimentary geothermal reservoirs can help infer the reservoir quality and improve the drilling success rate. The typical method for calculating the amount of quartz overgrowths involves a tedious process of manually examining Scanning Electron Microscope (SEM) images, particularly the cathodoluminescence (CL) and backscattered electron (BSE) images. This paper outlines an automated workflow to detect quartz overgrowths from SEM images using computer vision and machine learning techniques, which can dramatically cut the associated time and effort. The workflow was developed using SEM images available for public use. The automated workflow consists of a dynamic segmentation algorithm incorporating noise suppression, multi-level auto-thresholding, and dynamic overlaying. As a result, the workflow can automatically infer mineralogy from lower-quality images with varying brightness and contrast values and noise levels. The algorithm can also handle overlay shifting in CL and BSE images. Random Forest was used to train the algorithm based on extracted SEM image features such as Gabor, Canny Edge, and Roberts Edge. The resulting ML model is then used to improve the prediction of the image segmentation more accurately as a mineralogy predictor. The model training resulted in a 75% accuracy score, which is a promising start. The model can successfully differentiate between detrital quartz grains and their diagenetic quartz overgrowth; it can also identify porosity and the presence of other minerals. Furthermore, the detection capability was improved after training, particularly in reducing false positives in porosity detection. Further improvements can be made by applying morphology detection principles and expanding the model training to include different reservoirs. |