New Frontiers in Reservoir-Fluid Property Prediction: Redefining Accuracy Using Hybrid Artificial Intelligence Models and Extensive Datasets

Abstract This study presents an innovative artificial intelligence (AI)-based method that leverages abundant compositional analysis data—three to five times more available than full PVT studies—to predict critical reservoir fluid properties, specifically, bubble-point pressure, gas-oil ratio (GOR), API gravity, and formation volume factor. This approach reduces reliance on costly PVT experiments, enhances coverage and prediction accuracy, and improves reservoir fluid characterization, while incorporating advanced quality checks benchmarked against the entire dataset to identify outliers or experimental errors. The predictive model employs neural networks trained on PVT data, hybridized with empirical correlations and, optionally, physics-based models. In general, this integration yields more reliable predictions. To refine the model, network features are optimally selected, and rigorous quality checks are applied during both training and prediction phases. Increased robustness is achieved through an ensemble approach, analogous to bagging, in which multiple predictors are trained on distinct subsets of inputs. Data partitioning into subsets allows tailored training to improve prediction accuracy. The methodology was evaluated on real-field databases. Median predictions of fluid properties from 100 networks were utilized, showing notable improvements over uncalibrated equations of state, specifically the Soave-Redlich-Kwong and Peng-Robinson models, both with Péneloux volume correction. For the four properties analyzed, the mean absolute percentage errors (MAPEs) decreased by at least 20% compared to the equations of state, with values consistently below 10% for all properties except GOR. Bubble-point pressure was also estimated using ensemble networks enhanced at the input with Standing's correlation for that pressure. This enhancement reduced MAPE by 7% with respect to when only the networks are considered, outperforming the correlation alone by approximately 45%. Moreover, data partitioning resulted in a decrease of about 20% in MAPE for the prediction of bubble-point pressure and GOR compared to the non-partitioned model. This partitioning process involves evaluating a specific compositional-analysis input parameter, whose selection is determined through optimization. This approach strengthens fluid property estimation by hybridizing with empirical correlations to generate multiple predictions for probabilistic analysis, while increasing accuracy and robustness through data partitioning and quality checks. By requiring only compositional analysis, it reduces the need for expensive PVT laboratory experiments and also expands the spatial coverage of fluid data in reservoir characterization. This enables a more precise representation of reservoir fluid properties in the simulation model, ultimately improving operational efficiency and decision-making.