Physics-Informed Neural Networks for Two-phase Film Boiling Heat Transfer

In this paper, a physics-informed neural network (PINN) technique is developed to study a two-phase film boiling heat transfer process. Data generated through computational fluid dynamics (CFD) was used to train the PINN model. The formulated PINN approach was first validated against the classical Stefan phase-change study. Results show that the PINN predictions of interface location showed errors of up to 7.1% compared to the respective CFD solution. Subsequently, the PINN method was trained on a film boiling study with a Jakob number (𝐽𝑎 = 0.2). This PINN predictions in Nusselt number show a discrepancy of 6% compared to the CFD solution. Finally, the inference capabilities of the PINN approach were evaluated by applying transfer learning to predict the film boiling process with 𝐽𝑎 = 0.4 where no observational CFD data was provided (inverse problem). For this inverse case, the PINN predictions produced qualitative results which are in good agreement with unobserved reference data. Although small regions exhibited Nusselt number prediction errors of around 30%, it was found that these errors were predominantly caused by excessive interfacial diffusion. This study represents a groundbreaking development for PINN methodologies by applying the deep learning capabilities within to investigate the evolution of a film boiling process.