Hybrid framework provides high-resolution reconstruction of hypersonic wall-pressure signals

At hypersonic speeds, wall-pressure signals are integral to studying several performance-related parameters, including shock-boundary-layer interactions, aeroacoustic loading, and structural fatigue risk. Obtaining measurements in such extreme conditions to make high-fidelity simulations, however, is costly, requiring complex signals to be reconstructed from a handful of sparse measurements.

Fung et al. developed a reconstruction framework that uses AI to recover high-resolution hypersonic wall-pressure signals from sparse measurements. The framework combines a mathematical approach that creates smooth curves between data points, called cubic-spline interpolation, with a transformer model that further refines the reconstruction by learning spatial relationships that interpolation cannot capture.

Their work highlights the strengths and drawbacks of transformers, an AI architecture that is capable of learning elements of a system simultaneously and has powered recent advances in large language models.

“The division of labor between interpolation and the transformer reduces the burden on the learning model and makes the reconstruction more interpretable,” said author Dimitris Drikakis. “The work provides guidance on when broader non-local context is valuable, when local information is more reliable, and how sparse sensor layouts might be designed for hypersonic wall-pressure reconstruction.”

The group tested the hybrid framework’s performance across a range of parameters, including sparsity, sequence length, and turbulence regime, rather than treating accuracy as a single number. Such an analysis connects machine-learning design choices to flow-physics length scales.

The approach delivers mean absolute percentage errors less than 1% in minimally sparse setups and below 4% in moderately sparse setups. The researchers found that its accuracy is primarily affected by turbulence and input sequence length, with improved performance when spatial contexts capture correlation length scales.

The group next looks to analyze time-resolved and higher-dimensional reconstructions.

Source: “Uncertainty-aware artificial intelligence reconstruction of wall-pressure dynamics from sparse data in extreme flows,” by Daryl Fung, Nicholas Christakis, Ioannis William Kokkinakis, Dimitris Drika, S. Michael Spottswood, Kirk R. Brouwer, and Zachary B. Riley, Physics of Fluids (2026). The article can be accessed at https://doi.org/10.1063/5.0332704 .