Abstract

The energy transition involves decarbonization and the adoption of hydrogen and other renewable energy sources. Both CO₂ sequestration and hydrogen storage require subsurface rock formations with suitable porosity and permeability. The Combinable Magnetic Resonance (CMR) tool, a lithology-independent device, provides highresolution measurements of porosity and permeability, complementing conventional coring techniques. However, core permeability data is often unavailable due to technical issues such as poor recovery from unconsolidated sands, natural fractures, and borehole instability. In the study area, only a few wells have core data, and CMR logs are available for just one well. To address this limitation, supervised machine learning models are employed to predict Timur-Coates permeability (KTIM) and CMR-derived porosity (Magnetic Resonance Porosity, MRP) using basic wireline logs in the Barail Formation of the Assam Shelf Basin. Models considered include Random Forest (RF), Artificial Neural Networks (ANN), Linear Regression (LR), Support Vector Machine (SVM), Categorical Boosting (CATBoost or CATB), and Extreme Gradient Boosting (XGBoost). Among these, CATBoost showed the best performance in predicting porosity and permeability across depth. Hyperparameter tuning further enhanced its generalization ability. To assess the model's reliability, uncertainty quantification was also performed. The predicted values closely matched core measurements in permeable zones of the blind well.