Uncertainty-aware artificial intelligence reconstruction of wall-pressure dynamics from sparse data in extreme flows

Hypersonic environments constrain pressure instrumentation and make high-fidelity simulation datasets costly, leaving many applications with sparse wall-pressure measurements. This work develops and evaluates a hybrid reconstruction framework that recovers high-resolution hypersonic wall-pressure from strongly downsampled signals. The method first produces an initial estimate using cubic-spline interpolation, then refines it with a Transformer model that learns non-local spatial structure. Performance is assessed on datasets representative of Mach 5.3 wind tunnel conditions, spanning data-reduction factors from 2× to 64× and multiple turbulence-intensity regimes associated with shock–boundary-layer interaction. The hybrid approach delivers accurate reconstructions at low-to-moderate sparsity, with mean absolute percentage errors typically below 1% at low sparsity and generally below 4% at moderate sparsity, whereas degradation is primarily observed in the most extreme sparsity settings. A key finding is that reconstruction accuracy depends jointly on turbulence characteristics and input sequence length: a longer spatial context improves performance when the mean wall-pressure distribution exhibits longer correlation length scales, whereas a shorter context is preferable when the mean pressure exhibits more localized spatial variation (shorter correlation length scales). These results position Transformers as practical surrogates for reconstructing hypersonic pressure and provide actionable guidance on sensor sampling and sequence design in extreme-flow environments.