Exploring Physics-Informed Neural Networks for Compressible Flow Prediction

This work aims to understand the impact of hyperparameters of neural networks on the capability of physics-informed neural network (PINN) surrogates for compressible flow predictions and how they compare with traditional neural network (NN) surrogates by considering steady inviscid compressible flow in a 1-D converging–diverging nozzle subjected to different back pressures. Both NN and PINN are trained with known random flow characteristics estimated using analytical and computational methods corresponding to a set of back pressures and used to predict flow characteristics corresponding to arbitrary back pressures which are not in the training dataset.