Facies-Conditioned Physics-Guided Machine Learning for Interpretable Permeability Prediction in Heterogeneous Clastic Reservoirs

Abstract

Permeability prediction in clastic reservoirs is challenged by the limited vertical coverage of core measurements, the indirect nature of wireline log responses, and variations in pore connectivity across sandy and shaly lithological intervals. This study presents a novel facies-conditioned physics-guided machine learning workflow applied to the Zubair Formation in a well (TU-Well), Tuba oil field, southern Iraq. The final dataset comprises 178 depth-matched observations integrating core analysis, Nuclear Magnetic Resonance (NMR) logs, and conventional well logs. The dataset was split into 133 observations for training and 45 for a fixed development test. Initially, Random Forest, XGBoost, and CatBoost were evaluated as direct permeability models using well logs alone and then with NMR-derived permeability as an additional input. Subsequently, four independent operational clastic Electrofacies, Clean Sand, Shaly Sand, Transitional, and Shale, were identified from training-only gamma-ray and shale-volume data. The final architecture incorporates multi-resolution vertical context by sampling 10 continuous LAS curves within a ±5 m window around each core depth, generating 1,010 contextual values, which were then compressed into PCA20 and PCA60 representations for modeling. The final prediction is constructed as a hybrid model, combining an averaged CatBoost global context model with a strongly regularized soft-facies residual adapter, ensuring NMR permeability remains an explicit physical reference in the final reconstruction: K = KNMR+ ΔKglobal + ΔKfacies. While the best direct baseline model (CatBoost with logs and NMR) achieved an R2 of 0.6072 and RMSE of 197.46 mD, the final multi-resolution soft-facies approach yielded superior results: Random Forest achieved an R2 of 0.5978, XGBoost achieved 0.6561, and CatBoost achieved the highest performance with an R2 of 0.7265, RMSE of 164.76 mD, and MAE of 114.34 mD. The proposed hierarchy reduced RMSE by 16.56% relative to the best direct baseline and 50.90% relative to raw NMR estimates. The robustness of the approach was confirmed through three facies-stratified fourfold validations and one CORE_NO-grouped fourfold validation, where the hierarchy consistently improved RMSE across all 16 outer-fold comparisons.