Machine learning-driven permeability prediction in carbonates and sandstones using NMR relaxation data

Nuclear Magnetic Resonance (NMR) has proven to be a powerful tool for in-situ permeability quantification however, it typically requires laboratory calibration, and its accuracy is strongly influenced by rock type and pore system heterogeneity. Existing NMR-based permeability studies are often limited by small datasets, commonly restricted to a single lithology (sandstone or carbonate), and rarely investigate whether permeability prediction is more reliable using the full transverse relaxation time (T 2 ) distribution (spectrum) or NMR-derived parameters (e.g., minimum, maximum, peak T 2 …etc). As a result, the generalization of existing formulations across diverse geological settings remains limited. In this study, we address these gaps by developing machine learning models trained on a large and heterogeneous dataset of 308 core samples, including both sandstones and carbonates from the US, France, Middle East, and China. The dataset spans wide porosity (0.10–32.91%) and permeability (0.0003–15,400 mD) ranges, ensuring applicability across heterogeneous rock systems. Two algorithms, namely: Multilayer Perceptron (MLP) and eXtreme Gradient Boosting (XGB), are evaluated for predicting matrix permeability from NMR data. Model performance is compared using three input configurations: (i) NMR parameters extracted from the T 2 distribution, (ii) the full T 2 spectrum, and (iii) a combination of both extracted parameters and the full spectrum. This comparison assists in evaluating the added value of the full relaxation distribution, which captures pore-scale information that may be overlooked in simplified parameters. The results show that XGB consistently outperformed MLP, with the best performance achieved when combining the full T 2 spectrum and extracted parameters, yielding an R 2 of 0.86 and a root mean square error (RMSE) of 0.51 in log permeability prediction (corresponding to approximately 3 mD). Incorporating lithology (rock type: carbonate versus sandstone) as an input has only a minor effect on on XGB performance, suggesting prior lithological classification is not strictly required for accurate permeability prediction. These results indicate that the proposed approach can be generalized across sedimentary rocks and applied to both sandstone and carbonate reservoirs. • Developed machine learning models (XGB and MLP) to predict permeability from NMR T 2 relaxation data , addressing the limitations of traditional models like SDR. • Compiled a large, diverse dataset of 308 core samples (sandstones and carbonates) from multiple regions (U.S., France, Middle East, China), covering wide porosity (0.10–32.91%) and permeability (0.000296–15,400 mD) ranges. • Compared the predictive power of derived NMR parameters vs. the full T 2 spectrum , and tested the benefit of combining them for improved accuracy. • Demonstrated that XGB achieved the best results (R 2 = 0.86, RMSE = 0.51 in log permeability) when combining derived parameters with the full T 2 spectrum. • Showed that including rock type as an input had minimal effect , proving the model’s robustness and broad applicability across heterogeneous sandstone and carbonate systems.