Applying Green AI methods to Digital Rock Technology workflows

Digital Rock Technology comprises a set of image-based modelling techniques that analyse rocks at the pore level, gaining insight essential in industries including geological carbon dioxide sequestration, underground hydrogen storage, oil and gas, or geothermal energy. Pore Network Modelling simulates flow through a network of pores and throats representing a rock sample's void space, which results in improved efficiency and scalability compared to conventional simulation methods. We propose using semantic segmentation to identify the location and geometry of complex features such as fractures and vugs. Convolutional Neural Networks, a form of deep learning technology, can assist in automating the time-consuming process of handling extensive X-ray micro-computed tomography (micro-CT) data. //However, a solely accuracy-focused approach neglects the importance of efficiency and interpretability. This study explores the effect of deep-learning architecture choices on micro-CT semantic segmentation performance, considering computational resource efficiency and carbon footprint under the Green AI principles. The proposed novel objective function, incorporating topology and curvature information via Minkowski functionals, shows a 4% improvement in the accuracy of pre-trained models, and the study presents a strategy for selecting the optimal configuration. To address the lack of available labelled datasets, an algorithm was developed for generating semi-synthetic image-label pairs using features from publicly available data.