Predicting Bubble Point Pressure and Oil Formation Volume Factor by Explainable and Reproducible Support Vector Regression-Based Models from Multi-Regional PVT Data

Abstract

Accurate estimation of reservoir pressure-volume-temperature (PVT) properties underpins reserves evaluation, reservoir simulation, production forecasting, and surface facility design. Bubble point pressure (Pbp) and oil formation volume factor (Bob) are especially critical because they control hydrocarbon phase behaviour, fluid expansion, and depletion performance. Laboratory PVT analysis is the most reliable but often costly, slow, and unavailable. Empirical correlations are more practical; however, they suffer from regional bias, limited calibration ranges, and poor representation of nonlinear fluid behaviour. This study develops explainable, reproducible support vector regression (SVR) models to predict Pbp and Bob using four common descriptors: solution gas-oil ratio (Rs), gas specific gravity (γg), API oil gravity (γAPI), and reservoir temperature (TR). From 953 global PVT samples, 797 were preprocessed, normalised, and split using hold-out and k-fold cross-validation. Linear-kernel SVR delivered high accuracy: for Pbp, (R2 = 0.9826, R = 0.9913, RMSE = 0.0586, MSE = 0.0034); for Bob, (R2 = 0.9494, R = 0.9744, RMSE = 0.0882, MSE = 0.0078). To overcome the “black-box” limitation, the SVR models are fully specified using their support vectors, dual coefficients, and bias terms, enabling independent reproduction and implementation. Perturbation-based explainability shows that Rs is the dominant predictor for both properties, accounting for 71.87% of Pbp and 76.76% of Bob, consistent with thermodynamic expectations. Compared with widely used empirical correlations, the new models perform better on a multi-regional dataset. Thus, the explainable SVR models offer a transparent, accurate, and practical tool for reservoir PVT estimation, combining predictive performance with engineering interpretability.