A robust uncertainty quantification framework for machine learning–based wet-gas flow metering

Accurate uncertainty quantification (UQ) is essential for deploying machine learning (ML) models in multiphase flow metering, where limited training data, incomplete feature representation, and distribution shifts across operating conditions introduce significant epistemic uncertainty beyond the inherent variability of the sensors and flow dynamics. Using experimental datasets under diverse multiphase conditions, this study evaluates predictive uncertainty across five ML models, including deep neural networks (DNN), long short-term memory networks (LSTM), random forests (RF), extreme gradient boosting (XGBoost), and Gaussian process regression (GPR). Conformal prediction (CP) is employed as a model-agnostic framework to generate calibrated prediction intervals (PIs), whereas GPR estimates the predictive variance through its kernel-based structure. The evaluation results show that the gas flow rate predictions exhibit high accuracy and well-calibrated intervals across the models, with the CP producing fixed PIs and the GPR achieving the narrowest, dynamically adjusted intervals. Among the CP-based models, RF demonstrated the best balance between the maximum prediction accuracy and minimal uncertainty. Liquid flowrate predictions exhibited improved epistemic uncertainty across all models. In fact, introducing mixture fluid density, a potential engineered feature derived from the gas volume fraction (GVF) and phase densities, decreased uncertainties. The analysis continued using Explainable AI platforms to highlight the importance of features based on predictive strength. The study findings emphasize the importance of both targeted feature engineering and uncertainty-aware modeling, highlighting practical advantages of CP for model-agnostic UQ and the necessity of XAI tools for transparency. Overall, these results support the applicability of explainable, uncertainty-calibrated ML systems for real-time multiphase flow monitoring, with direct implications for metering confidence and operational decision-making.