NMR-data-driven prediction of matrix permeability in sandstone aquifers

•Matrix k was accurately predicted by machine learning models from the NMR T2 data.•The cumulative T2 data as inputs provided higher accuracy for k than original bins.•The optimal GBDT model was visualized and interpreted by the game-theory SHAP method.•The classical SDR and Timur-Coates models underestimated k by the fixed constants.•This work provides insight into the predictions of ‘from curves to values’. Predicting the matrix permeability of subsurface sandstone aquifers is a formidable challenge. One test showing great promise is nuclear magnetic resonance (NMR), as it is the only advanced hydrogeophysical technique non-invasively measuring subsurface pore structure information and being applied in downhole well-log analysis with core calibration. Unfortunately, the classical NMR-based permeability models suffer from more than two empirical constants that limit prediction accuracy greatly. In this study, a NMR-data-driven method was reported with nine machine learning models. Results indicate that the type of input data strongly influences the machine learning modeling of matrix permeability. Using the cumulative T2 relaxation data instead of the original T2 data, a stronger agreement was found between experimental results and NMR estimates of matrix permeability from the gradient boosting decision tree model automatically tuned by GridsearchCV. The correlation coefficient reaches 0.92 with the lowest MSE of 0.12. Due to the nature of “black box” of machine learning model, the visualization and interpretation of the optimal model were performed through the Shapley Additive exPlanations method. The 38th variable of 501 ms in the cumulative T2 spectrum positively contributes the most to the matrix permeability. However, the 26th variable of 31.6 ms provides the largest negative contribution to the final predicted matrix permeability. This work aims to provide a new idea to the “from curve-to-value” prediction process for a rapid and accurate matrix/bedrock permeability estimation via NMR T2 data in deep sandstone aquifers and extensive porous rocks.