Autoencoder-Based RBF and LSTM for Extrapolation Prediction of Hypersonic Inlet Unsteady Flow

Prediction of the hypersonic inlet unsteady flowfield is crucial for preventing inlet unstart. By combining an autoencoder (AE) for dimensionality reduction with radial basis function (RBF) methods and long short-term memory (LSTM) networks, models for predicting the hypersonic inlet unsteady flowfield were constructed. This study compared the interpolation and extrapolation performance of these models, including several AE-RBF variants with different RBF types and shape parameters, as well as the AE-LSTM model. The results show that within the sample space, the cubic RBF method can accurately predict the unsteady flowfield, but it fails rapidly outside the sample space. The Gaussian RBF yields relatively stable prediction errors with negligible influence from the shape parameter, achieving an extrapolation prediction error of 5.7% at 70 ms after backpressure was applied. The extrapolation error of the multiquadric RBF increases with time, and a larger shape parameter leads to a greater error. Although the extrapolation error of the AE-LSTM method also rises slowly with time, it remains as low as approximately 2.01% at 70 ms post-backpressure. This study reveals that the Gaussian AE-RBF method and the AE-LSTM method have good potential for predicting the unsteady flowfield of the hypersonic inlet.