A hybrid deep learning framework for the prediction of variable-sweep aircraft surface pressure distribution

The rapid and accurate prediction of surface pressure distributions on variable-sweep aircraft remains a significant challenge, particularly during the transient morphing process. This paper presents a novel hybrid deep-learning framework to address this issue by synergistically integrating a convolutional autoencoder (CAE), a multilayer perceptron (MLP), and a long short-term memory (LSTM) network. The core of our methodology is a dimensionality reduction via a CAE, which non-linearly compresses high-dimensional (384 × 128) surface pressure fields into a 12-dimensional latent space, achieving a 4096:1 compression ratio with a reconstruction mean squared error (MSE) of 3.5074 × 10−5. Subsequently, the prediction task is decomposed: an MLP network accurately maps flight parameters to latent variables for steady-state predictions, achieving a correlation coefficient of R > 0.99, while an LSTM forecasts the temporal evolution of the latent space for unsteady morphing sequences with a mean absolute error below 5.0 × 10−3. Upon decoding, the fully reconstructed surface pressure fields exhibit an overall MSE below 7.6 × 10−5 for steady cases and 7.5 × 10−5 for the dynamic variable-sweep process. Results demonstrate that the proposed framework achieves a high predictive accuracy for both static configurations and dynamic morphing sequences, while offering a substantial computational speed-up compared to conventional computational fluid dynamics solvers. This work establishes a robust and efficient data-driven paradigm for aerodynamic analysis, significantly accelerating the design and optimization cycle for morphing aircraft.