Physics-Informed Explainable Machine Learning for Tight Sandstone Facies Prediction from Conventional Well Logs: A Blind-Well Comparative Study Krishna-Godavari Basin, INDIA

Abstract Background Conventional hydrocarbon exploration worldwide with India as not exceptional is increasingly finding smaller, deeper and more marginal accumulations , the transition from "find another conventional field" to target "productivity from tight reservoirs" could become one of the most important exploration-development strategies over the next decade along with deepwater exploration. Tight sandstone reservoirs commonly exhibit complex lithological and petrophysical characteristics and require stimulation for production. Reliable identification of tight sandstone facies from conventional well logs is challenging. Although machine-learning techniques have increasingly been applied to automated lithofacies classification, most of the studies rely on random sample-wise validation, which may overestimate predictive performance because adjacent well-log samples are highly correlated. Furthermore, the incremental value of physically derived petrophysical, petro-elastic and geo-mechanical attributes over conventional log measurements remains insufficiently quantified. Objective This study develops an explainable machine-learning framework for four-class lithofacies classification of deep-seated syn-rift sediments of KG Basin comprising primarily Shale, Siltstone, Clean Sandstone and Tight Sandstone—and investigates whether computed petrophysical and petro-elastic attributes improve tight sandstone discrimination along with identification of zones which would require stimulation and well-to-well generalization. Methodology Three wells with interpreted lithofacies are used for model development and blind-well validation, while two additional wells are reserved for independent prediction. Three principal feature configurations are evaluated: (i) conventional well logs and (ii) conventional logs augmented with computed petrophysical, petro-elastic and geomechanicalattributes . Random Forest, Catboost, LGBM and XGBoost classifiers are evaluated using reciprocal well-wise validation with same prediction parameters. (iii) Model performance is assessed using accuracy, balanced accuracy, macro-F1 and class-specific precision, recall and F1 scores, with particular emphasis on Tight Sandstone classification . SHAP-based explainable AI is employed to identify the dominant geological and elastic attributes controlling model predictions. Prediction probabilities and model disagreement are used to characterize uncertainty. Expected contribution The study establishes whether physics-informed feature engineering provides measurable improvement in blind-well lithofacies prediction and determines which conventional and derived properties contribute most strongly to Tight Sandstone identification. The resulting workflow provides an interpretable framework for extending facies interpretation into wells lacking direct lithofacies information. Results The study shows that ensembles of machine-learning classifiers can accurately classify Shale, Siltstone, Clean Sandstone, and Tight Sandstone from conventional well logs, with physics-informed petro-elastic attributes significantly improving accuracy. Overall, using extended attributes increased mean accuracy by 14–16% and mean ROC-AUC by approximately 9% and resulted in ΔF1 of 0.1869 (30.86% improvement and 95% bootstrap CI: 0.1766–0.1966). For Tight Sandstone, the Random Forest classifier achieved a precision of 0.979, a recall of 0.996, and an F1-score of 0.987. The study used SHAP analysis to identify physically informed petrophysical and elastic attributes that contributed to the model’s predictions and used prediction probabilities and model agreement/disagreement to highlight uncertainty. The study recommends applying the well-log–petro-elastic–explainable ML framework for analyzing tight-sandstone deposits and using blind-well validation to assess the method’s transferability and avoid overestimated performance metrics.