Shear wave velocity prediction from minimal well logs using neighborhood component analysis—guided XGBoost in carbonate reservoirs

Shear wave velocity (Vₛ) is a critical parameter for petroleum geomechanics and reservoir engineering. Direct measurement via logging tools is often restricted due to economic and technical constraints, necessitating reliable indirect prediction methods. Traditional empirical correlations are lithology-dependent and lack generalization, while many machine learning (ML) approaches suffer from overfitting and limited interpretability due to suboptimal feature selection. This study introduces an interpretable computational framework integrating Neighborhood Component Analysis (NCA) with ML algorithms for accurate Vₛ estimation. Six ML models: XGBoost, Random Forest, Decision Tree, Support Vector Regression, K-Nearest Neighbors, and Linear Regression, were evaluated using well-log data from a carbonate reservoir. The NCA-XGBoost model achieved superior accuracy (R 2 = 0.946, RMSE = 0.068 km/s), significantly outperforming the best empirical correlation (Castagna and Backus, R 2 = 0.698, RMSE = 0.075 km/s), representing an approximately 35% improvement in R 2 and a 10% reduction in RMSE. NCA identified compressional wave transit time (DTCO) as the dominant predictor, with a single log achieving near-optimal accuracy. This approach reduces dimensionality, mitigates overfitting, and enhances model transparency by identifying the dominant physical controls on Vₛ and quantifying their relative importance. The proposed framework offers a cost-effective, reliable alternative for Vₛ prediction in wells lacking direct measurements, with practical implications for wellbore stability analysis, hydraulic fracturing design, and seismic interpretation.