A Multi‐Model, Physics‐Guided Machine Learning Approach for Lithofacies and Petrophysical Prediction in Complex Carbonate Reservoirs of the West Coast of India

Abstract

Accurate lithofacies and petrophysical prediction in carbonate reservoirs remains challenging due to substantial heterogeneity, nonlinear rockfluid interactions and well log acquisition biases. This study presents a physics‐guided machine learning workflow that integrates domain‐driven feature engineering, Shapley Additive exPlanations (SHAP), baseline alignment and blind well validation to improve lithological and petrophysical characterization in complex carbonate formations. Conventional wireline logs, along with petrophysically meaningful derived features, were used to capture density–porosity–sonic–resistivity relationships. Four advanced regression models, Multilayer Perceptron Regressor (MLPR), Extreme Gradient Boosting Regressor (XGBR), Light Gradient Boosting Machine Regressor (LGBMR) and CatBoost Regressor (CBR), were trained to predict fractional lithologies and key petrophysical parameters, including shale volume, multiple porosity components and water saturation (SW). SHAP analysis was employed to verify that model predictions are controlled by physically consistent inputs rather than spurious correlations, whereas systematic baseline shifting aligned blind well logs with the training domain to remove acquisition‐related bias. Results show that predictions for limestone and shale are primarily governed by density, porosity, sonic, resistivity and caliper logs, whereas volumetric shale is effectively controlled by gamma ray attributes. Porosity components exhibit strong inverse relationships with density and consistent sensitivity to sonic response, and SW is primarily driven by resistivity, confirming petrophysical consistency across all models. Blind well predictions demonstrate stable generalization, with MLPR producing smoother, geologically coherent trends, whereas tree‐based models capture sharper transitions and localized heterogeneity. Lastly, this study shows that no single model fully captures the multiscale variability of carbonate systems. Instead, a physics‐guided multi‐model strategy provides the most robust representation of subsurface properties. The proposed workflow provides a transparent and rock physics‐consistent framework for carbonate reservoir characterization, with model predictions controlled by physically meaningful well log relationships, validated through SHAP analysis.