Velocity Prediction in Hypersonic Flow via Prior-Guided Data Fusion and Physics-Informed Learning

Abstract

Accurate velocity prediction of bodies under hypersonic flow is crucial for aerospace engineering and safety evaluation. However, unsteady aerodynamic effects coupled with sparse and noisy experimental data make accurate velocity prediction challenging. Taking conical-end cylinder as example, a prior-guided data fusion method based on physics-informed learning is proposed. First, flight simulations for bodies under multiple conditions are performed to construct a quasi-steady drag prior model. The prior model is embedded into the PINN, where sparse experimental velocity data are used for supervision, while Newton’s law of motion is incorporated as a physical constraint. By learning a time-dependent correction factor for the quasi-steady drag coefficient, the proposed method reduces the differences between the velocity prediction based on the quasi-steady aerodynamic model and experimental measurements. The proposed method yields velocity prediction errors of 0.97% and 0.41% under typical conditions of Mach 5.98 and 6.62, respectively. Under extrapolation conditions deviating by 15% from the training range, the velocity prediction error remains below 1.74%. Compared with neural network methods based on quasi-steady simulation results, as well as Support Vector Regression and Random Forest methods based on experimental data, the proposed method demonstrates higher prediction accuracy and improved extrapolation capability for velocity prediction under hypersonic flow conditions.