Physics-informed neural networks to predict mass transfer at interfaces in liquid-vapor phase change problems

This work demonstrates the use of Physics-informed Neural Networks (PiNNs) to infer the continuous flow fields and local mass transfer in the liquid-vapor phase-change problem of a vapor bubble rising in a subcooled liquid domain. The model does not assume any mass transfer model and uses time-scattered data of the volume fraction and liquid velocity to predict the flow fields and the mass transfer. A synthetic dataset was generated using computational fluid dynamics (CFD) simulations, considering an empirical heat and mass exchange model for mass transfer, and the relevant parameters at time-scattered values were extracted from these simulations. A PiNN model was trained with the governing physics-based partial differential equations (PDEs) and the observed data as the loss functions to infer the continuous mass transfer. Dynamic weighting and residual-based pointwise attention were implemented in the PiNN model to improve the accuracy of the predictions. The results show that local mass transfer can be inferred from CFD data, which is a substitute for experimentally observable data for the present work, combined with the governing PDEs, without presupposing any mass transfer model, paving the way for extracting mass transfer for more complex cases to improve the existing mass transfer models.