Lithology-Guided Shear Wave Velocity Prediction: Ensemble Machine Learning and Composite Models for Bridging the Generalization Gap

A very important parameter in seismic reservoir characterization and geomechanical modeling is the shear wave velocity (Vs). Nevertheless, it is not easy to quantify in the wells because the acquisition cost can be very expensive, even though machine learning (ML) is an affordable method for creating synthetic logs; random train-test splitting is common in traditional research. This approach does not consider spatial heterogeneity, leading to excessively optimistic projections of generalization. To explore this critical gap, this paper develops a lithology-sensitive composite model of Vs prediction, and the model is thoroughly tested on a blind-well dataset of complex lithologies of sandstone, heterolithic, and shales. Three different machine learning models (e.g., Random Forest (RF), Artificial Neural Network (ANN), and Extreme Gradient Boosting (XGB)) and composite models are evaluated to determine their predictive effectiveness, as well as the existing empirical correlations, in the case of domain-shift. The findings showed that there is no dominant ML algorithm in all lithological facies. Sandstone, heterolithic, and shales are used with the algorithms of RF, ANN, and XGB, respectively. Results revealed that the composite model worked the best with an operational precision of 71.0 percent of predictions within the ± 150 m/s of measured values, which were much better than individual ML models and empirical correlations. This paper indicates that the ensemble-based geologic methodology could offer a sound methodology in forecasting Vs of basins that have complex geology and scanty information.