Cluster-Specific Random Forest Models for Predicting Compressive and Shear Slowness from Conventional Logs in the Bassein Formation of Western Offshore India

ABSTRACT Shear and compressional wave velocities are fundamental to reservoir lithology, rock mechanics, and fluid characterisation in geophysical exploration. However, shear wave logs are often absent due to limitations in the tool or high acquisition costs. This study presents a machine learning workflow to predict compressional (DTC) and shear (DTS) travel times from conventional well logs in the hydrocarbon-bearing Bassein Formation, located off the Western Coast of India. The carbonate system is strongly influenced by variations in porosity, diagenetic alterations, and facies heterogeneity, making velocity prediction particularly challenging. The workflow integrates unsupervised clustering with supervised random forest regression. Clustering partitioned the dataset into three groups representing distinct lithological and petrophysical conditions, enhancing model sensitivity to heterogeneity. Model performance was assessed using the coefficient of determination (R2) and Root Mean Square Error (RMSE). Cluster 1, representing more homogeneous carbonate facies, achieved the best balance between variance explanation and error magnitude (training R2 = 0.9223, RMSE = 4.46; validation R2 = 0.7524, RMSE = 7.79). Cluster 2 showed slightly higher training R2 (0.9334) but a larger error (RMSE = 4.79), while Cluster 3 obtained the highest training R2 (0.9417) but the poorest error control (RMSE = 6.51; validation RMSE = 11.86). Cluster 1 proved to be the most reliable as it combines strong predictive power with controlled error, unlike Cluster 3, where greater variance capture came at the cost of larger prediction errors. Weighted performance across all clusters yielded a training R2 of 0.9291 (RMSE = 4.77) and a validation R2 of 0.7731 (RMSE = 8.47), confirming robust predictive capability for the dataset as a whole. The results highlight that facies heterogeneity, porosity distribution, and diagenetic features strongly control machine learning accuracy. By integrating geological insight with data-driven methods, this study establishes a reliable framework for reconstructing missing sonic logs and improving reservoir characterisation in complex carbonate systems.