A Locally enhanced physics-informed neural network surrogate model for rapid aerothermal assessment of aero-engine nozzles

Abstract

A novel deep learning surrogate model based on a Locally Enhanced Physics-Informed Neural Network (LE-PINNs) is proposed to address the challenge of rapid and accurate prediction of aerothermal and aerodynamic performance for aero-engine exhaust nozzles under varying operating conditions. The model takes geometric information and inflow boundary conditions as inputs, outputting flow field variables (density, pressure, temperature, etc.). The core innovation lies in its unique dual-network collaborative prediction framework: the global network, constrained by the Reynolds-Averaged Navier-Stokes (RANS) equations, captures primary flow field characteristics, while the boundary network specializes in high-precision modeling of pressure and temperature distributions near the nozzle walls. The model was trained and validated using a high-fidelity dataset generated by high-precision CFD simulations, covering a range of nozzle pressure ratios (NPR) from 4.0 to 14.5 and corresponding variations in the expansion area ratio (AR).

Test results demonstrate that, LE-PINNs exhibits superior prediction accuracy across the entire flow field, particularly near the walls, achieving an average relative error in wall temperature predictions below 1 %, significantly outperforming than Fully Connected Neural Networks (FCNN) and standard Physics-Informed Neural Networks (PINNs) models. In predicting key parameters at the nozzle outlet (temperature, velocity, thrust, etc.), LE-PINNs maintains deviations below 1 % compared to CFD results, accurately capturing shock wave structures and their intensities. Furthermore, its prediction efficiency is approximately 360 times faster than traditional CFD methods for single-case inference within the trained parameter range (NPR from 4.0 to 14.5 and AR from 1.22 to 2.39).