Filling-ratio-dependent heat transfer behavior and machine-learning prediction of hybrid water–paraffin phase change thermal storage

This study investigates interfacial replacement-controlled melting in an immiscible water–paraffin thermal storage system. Unlike conventional phase change storage, in which heat transfer is mainly limited by the low conductivity of phase change material (PCM) and the increasing distance between the heat source and the retreating melting front, the present configuration introduces water as a mobile heat transfer phase above the paraffin layer. This phase redistribution creates a moving thermal pathway that sweeps the solid–liquid interface, weakens the late-stage thermal resistance, and suppresses the formation of low-temperature refractory zones. A coupled VOF-based numerical model validated by visualization and temperature measurements is used to resolve the evolution of temperature fields, liquid fraction, phase distribution, and heat storage rate under different water–paraffin filling ratios. Increasing the water fraction from 20% to 80% reduces the paraffin melting time by 76.40%, primarily due to intensified natural convection and weakened interfacial thermal resistance. Meanwhile, the total heat storage capacity decreases by 22.06%, indicating an inherent trade-off between heat transfer enhancement and energy storage density. Dimensionless analysis confirms the dominant transport regime. The Rayleigh number reaches the order of 10 8 –10 9 , demonstrating convection-dominated melting, while the Stefan number remains approximately 0.1, revealing that latent heat absorption continues to dominate the overall energy storage process. The Nusselt number is on the order of 10 2 , further verifying the substantial enhancement of convective heat transfer relative to conduction-controlled PCM melting. A data-driven surrogate model is further developed as a secondary tool for rapid prediction of system-level thermal indicators after the CFD-based mechanism analysis. Long Short-Term Memory (LSTM) model achieves prediction errors of 1.05% for paraffin heat storage and 1.52% for water related thermal metrics, whereas the backpropagation (BP) model maintains errors below 0.6% under smoother thermal response conditions. These findings clarify the coupled effects of filling ratio, natural convection, and interfacial heat transfer on hybrid phase change melting, and provide a data-driven approach for rapid prediction of multiphase transient heat transfer processes.