Robust physics-informed neural network and swarm-based modeling of subcooled flow boiling pressure drop in metal foam tubes

Purpose To develop a hybrid machine-learning framework that integrates a Physics-Informed Neural Network (PINN) with Particle Swarm Optimization (PSO)–tuned regression models to deliver fast, accurate, and physically consistent pressure-drop predictions in metal-foam boiling applications. Design/methodology/approach • A novel PINN is constructed by embedding the full set of governing thermal–hydraulic equations (mass, momentum, and energy conservation) directly into its loss function. • Five regression models (SVR, AdaBoost, Polynomial Regression, XGBoost and Random Forest) are hyperparameter-tuned via PSO. • Both the PSO-optimized regressors and the PINN are trained and tested on synthetic datasets representative of subcooled boiling flows in metal-foam tubes to assess accuracy, robustness, and computational efficiency. Findings The PINN model outperforms all baseline PSO-tuned regressors, achieving an R2 of 0.9947 on test data. This demonstrates that enforcing physical laws during training yields superior generalization and predictive performance compared with purely data-driven approaches. Research limitations/implications - Validated on synthetic and experimental data, providing practical confidence.- Extensible to diverse foam geometries and fluids for scalable applicability.- Compatible with sensor networks and real-time control for accurate, responsive pressure-drop forecasts. Practical implications High precision RF, PSO tuned XGBoost and PINN models integrate into digital twins and control loops for dynamic pressure drop prediction, anomaly detection, and adaptive setpoint adjustment. This reduces downtime, optimizes heat exchanger design, lowers maintenance costs, and enhances system reliability. Social implications Enables safer, more efficient boiling systems with less downtime and environmental impact. Originality/value This work uniquely combines PSO-optimized ensemble learners with a physics-driven PINN architecture that explicitly embeds conservation laws into the learning process, an innovation not previously applied to subcooled boiling in porous media. It establishes a new benchmark for constraint-aware, data-driven modeling in thermal-hydraulic research.