A data-driven phase change model for injection flow modeling

A deep learning approach is developed to swiftly evaluate phase change in computational fluid dynamics (CFD) simulations of a multi-component, liquid–gas two-phase injection flow. This method significantly improves computational efficiency by using a deep feedforward neural network (DFNN) to replace the complex iterative solution of multi-species vapor–liquid equilibrium (VLE). The DFNN takes instantaneous pressure, temperature, and system composition as input and predicts the corresponding phase equilibrium state. A parametric study was conducted to optimize the neural network's hyperparameters, including the activation function, number of hidden layers, and neurons per hidden layer. The rate of phase change is then calculated as a linear relaxation toward phase equilibrium, guiding subsequent computational steps in the CFD solver. A case study was performed to test the proposed methodology, involving the injection of a superheated liquid ethanol–water mixture into a gaseous nitrogen environment. The simulation results and computational cost were examined. It is found that the DFNN model, while accurately representing the non-ideal non-equilibrium phase change of a multi-component injection flow, speeds up the VLE solution by four orders of magnitude, leading to a 30%–40% reduction in overall flow simulation time. This model shows promise for injection flow simulations, especially for systems with a large number of compositions, such as sustainable aviation fuels.