Evaluation of Deep Learning Architectures for Local Flow Parameter Estimation in Gas–Liquid Two-Phase Flow Under Variable Conditions

Abstract

Gas–liquid two-phase flows arise in pipelines, power plants, and process equipment where reliable prediction of local flow parameters is required for design, control, and safety. Local quantities such as phase volume fraction, phase velocity, and interfacial area are strongly affected by flow regime transitions and operating conditions, which makes empirical correlations difficult to generalize. High-fidelity computational fluid dynamics can resolve these fields, but the computational cost prohibits their use in real-time applications or large parametric studies. Recent progress in deep learning offers alternatives for learning low-cost surrogates directly from data collected by numerical simulation and experimental sensors. This work evaluates the ability of different deep learning architectures to infer local flow parameters from available measurements under variable gas and liquid flow rates, fluid properties, and pipe inclinations. Convolutional, recurrent, and attention-based models are trained on synthetically generated and experimentally inspired data that span stratified, slug, and annular flow conditions. The study compares prediction accuracy, robustness to measurement noise, and generalization to unseen operating conditions, and examines how architectural choices affect the reconstruction of small-scale interfacial features. Furthermore, the analysis explores the benefits of incorporating approximate physical constraints during training. The results highlight consistent trends in the relative performance of competing architectures and map regions in the operating space where data-driven estimators remain reliable, providing guidance for selecting models in practical monitoring and control systems.