Direct Prediction of Oil Saturation from NMR T₂ Spectra Using Unsupervised Feature Decoupling and Deep Learning

Nuclear magnetic resonance (NMR) T₂ spectra provide pore-scale information on fluid occurrence and mobility and are widely used for saturation evaluation. However, in shale and other unconventional reservoirs, strong heterogeneity, multi-fluid signal overlap, and weak/complex relaxation responses often undermine the reliability of conventional cutoff- and template-based saturation methods.

To address this challenge, we propose a data-driven workflow that directly predicts oil saturation from NMR T₂ spectra by integrating feature engineering, unsupervised decoupling, and a compact deep-learning regressor. The method first applies robust preprocessing to suppress abnormal values and outliers, followed by dimensionality reduction to extract the most informative latent features. To alleviate multi-component signal superposition, an unsupervised clustering step is introduced to partition spectral patterns into representative groups, providing a more stable feature basis for learning. Finally, a lightweight convolutional neural network (CNN) is employed as the regression model to map processed T₂ features to core-calibrated oil saturation, with standard strategies (normalization, dropout/regularization, and learning-rate scheduling) to improve generalization.

The workflow is validated using core-log paired datasets from a shale reservoir, showing that the predicted oil saturation agrees well with laboratory measurements and significantly improves stability compared with conventional interpretation in complex intervals. The proposed approach offers an efficient and scalable route for saturation evaluation in data-limited unconventional plays, supporting sweet-spot identification and development planning.

This research was supported by the National Oil & Gas Major Project (No. 2025ZD1400202) and Natural Science Foundation of Shandong Province, China (No. ZR2023YQ034).