Abstract
Nations relying on nuclear power generation face great responsibilities when designing their firmly secured final repositories. In Hungary, the potential host rock [the Boda Claystone Formation (BCF)] of the deep geological repository is under extensive examination. To promote a deeper comprehension of potential radioactive isotope transport and ultimately synthesis for site evaluation purposes, we have efficiently tailored geospatial image processing using a convolutional neural network (CNN). We customized the CNN according to the intricate nature of the fracture geometries in the BCF, enabling the recognition process to be particularly sensitive to details and to interpret them in the correct tectonic context. Furthermore, we set the highest processing scale standards to measure the performance of our model, and the testing circumstances intentionally involved various technological and geological hindrances. Our presented model reached ~ 0.85 precision, ~ 0.89 recall, an ~ 0.87 F1 score, and a ~ 2° mean error regarding dip value extraction. With the combination of a CNN and geospatial methodology, we present the description, performance, and limits of a fully automated workflow for extracting BCF fractures and their dipping data from scanned cores.