Search for a command to run...
Abstract Accurate estimation of Poisson's ratio (ν) is a critical requirement for wellbore-stability analysis, in-situ-stress estimation, drilling-performance optimization, and hydraulic-fracturing design. Conventional determination of ν relies on laboratory measurements or empirical correlations derived from sonic logs, which are often costly, time-consuming, and inadequate for capturing the complex geomechanical behavior of heterogeneous clastic reservoirs. Traditional empirical models are overly simplified and fail to account for lithological variability, while purely data-driven machine learning (ML) approaches, although powerful, lack physical consistency and interpretability, leading to poor generalization outside the training domain. This study introduces an explainable Physics-Guided Machine Learning (PGML) framework that integrates domain-specific geomechanical constraints with advanced machine-learning models capable of improving the prediction of Poisson's ratio (ν). The methodology is based on discrepancy modeling, where an empirical physics-based equation is first used to estimate ν, and machine learning models are then trained to learn and correct the difference between the empirical estimate and laboratory-measured values. Five machine learning algorithms were implemented and compared in both conventional ML and PGML modes: Adaptive Boosting (AdaBoost), Gradient Boosting (GB), Random Forest (RF), Categorical Boosting (CatBoost), and Radial Basis Function Neural Network (RBFNN). The models were trained using conventional well logs, including porosity, bulk density, gamma ray (GR), deep resistivity, slowness (ΔT), ultrasonic P-wave velocity (Vp), depth, and lithofacies. The PGML approach consistently outperformed conventional ML across all models. The AdaBoost-PGML model achieved the highest predictive performance, improving R2 from 0.836 to 0.975 and reducing RMSE from 0.0193 to 0.0075. Similar performance gains were observed for RF, CatBoost, and RBFNN models, demonstrating the robustness of the PGML framework. To enhance transparency, SHapley Additive exPlanations (SHAP) were applied to identify the relative influence of each input feature. Bulk density, porosity, gamma ray, and depth were found to be the most influential parameters contributing to the discrepancy between empirical estimates and measured ν values. The proposed PGML framework provides a physically consistent, interpretable, and data-driven approach for real-time prediction of Poisson's ratio based on readily available well-log inputs. This hybrid methodology reduces reliance on expensive laboratory testing, improves prediction reliability in heterogeneous formations, and supports more informed geomechanical decision-making in reservoir engineering and drilling operations.