Applications of self-supervised deep learning in micro-tomography of pre-salt carbonate rocks

Resumo

Characterizing rocks, especially determining petrophysical properties of reservoir rocks, remains a key challenge in the petroleum industry. Techniques like numerical modeling simulate reservoir behavior in porous media but depend critically on accurately assessing textural characteristics.CT and micro-CT are widely used to analyze rock samples, while digital image analysis quantifies grain and pore features in heterogeneous rocks like carbonates. Pore space and mineralogy aid in selecting representative samples, though integrating geological parameters into digital models remains essential. The complexity of carbonate rocks continues to pose challenges.Deep learning, leveraging multi-layered neural networks, excels in extracting patterns from data without extensive preprocessing, enabling automation of rock image analysis and property estimation. A major challenge lies in maximizing unannotated data usage while learning from limited annotated samples. Self-supervised frameworks address this by pre-training models on pretext tasks before fine-tuning for downstream applications like permeability estimation. This approach reduces reliance on annotated data, leveraging abundant unannotated samples effectively.This work proposes the use of self-supervised deep learning models for different predictive tasks: image recovery, absolute permeability estimation, and porosity estimation. The permeability/porosity values for each sample were obtained through laboratory tests. The image recovery task was chosen because rock micro-CT images contain observable visual patterns, making it a feasible task. Furthermore, few open rock datasets are available. No large annotated carbonate micro-CT dataset exists, primarily because access to such data is restricted to companies. Additionally, based on a literature review, this is the first study to apply self-supervised deep learning models to support reservoir characterization using pre-salt carbonate rock micro-CT images.