Informer deep learning forecasting for high-speed aeroelastics

Abstract

This paper examines the suitability and limitations of the Informer deep-learning architecture for forecasting aeroelastic response in high-speed shock-wave/turbulent boundary-layer interaction (SBLI). Dynamic displacement measurements of a thin, compliant panel were obtained in the H2K hypersonic wind tunnel at Mach 5.3 for several cavity-pressure configurations. The Informer, which employs ProbSparse self-attention for long-sequence time-series forecasting, was trained on displacement time histories from a subset of cases and used to predict unobserved scenarios. We systematically varied the forecast horizon (256-768 samples) and data sparsity (1200-4000) and evaluated performance using the normalised root-mean-square error. The model achieved low error (minimum NRMSE 0.19) only for carefully tuned combinations of step size and sparsity, while many other configurations yielded NRMSE values of O(0.5-1) and exhibited systematic under- or overestimation. In particular, the Informer tended to underestimate displacement following recent sharp decays and to overestimate signals that contained an initial near-steady phase. These behaviours reveal a strong sensitivity to hyperparameters and difficulty in capturing the non-stationary, thermally buckled, shock-induced dynamics of the panel. The results indicate that, in its standard form, the Informer is not robust enough to serve as a stand-alone predictor for aeroelastic design. Still, it provides a valuable diagnostic benchmark for long-sequence attention models in hypersonic fluid-structure interaction. The study highlights the need for adaptive or hybrid architectures and suggests future applications in virtual sensing and safety monitoring for high-cost hypersonic testing.