High-Speed Fluid-Structure Interaction Predictions using Deep Learning Transformer Architecture
Professor Dimitris Drikakis, Daryl Fung, Dr. Ioannis Kokkinakis, University of Nicosia (UNIC),
S. Michael Spottswood, Kirk R. Brouwer, Zachary B. Riley Air Force Research Laboratory (AFRL),
Dennis Daub, and Ali Gülhan, German Aerospace Center (DLR)
In high-speed aerospace vehicles, the interaction between shock waves and boundary layers can lead to significant structural deformations and premature failure. While experimental and computational methods exist to study these phenomena, they are often expensive and resource-intensive. This research introduces an innovative approach using artificial intelligence to predict fluid-structure interactions in hypersonic flows.
This paper presents the development and application of a Transformer deep-learning model to predict fluid-structure interactions induced by shock-turbulent boundary layer interaction. The model was trained using experimental data from a hypersonic wind tunnel, operating at Mach 5.3 and Reynolds number of ~19.3×106/m.
The paper was published on AIP Physics of Fluids on 12 May 2025 and was selected as an Editor’s Pick.
Objectives of the study
What Does This Mean?
- The Transformer model can predict aeroelastic panel deformations using only cavity pressure measurements.
- The approach demonstrates the potential for integrating AI methods with aerospace experiments, potentially reducing costs and resource requirements.
High-Speed Fluid-Structure Interaction Predictions using Deep Learning Transformer Architecture
Read the full paper
Non-Technical Summary
The research introduces an artificial intelligence approach to predict how structures behave under hypersonic flow conditions. Using a type of AI called a Transformer, the study shows that it’s possible to predict how panels deform under extreme conditions by only measuring the pressure in a cavity beneath the panel. The model was tested using data from experiments conducted in a hypersonic wind tunnel and showed promising results, particularly when trained with more data samples. This approach could significantly reduce the need for expensive experiments and complex computations in aerospace design.
Contact Information
For more information about this study, please contact Professor Dimitris Drikakis, Dean of the UNIC School of Sciences and Engineering, UNIC Vice President for Global Partnerships and Executive Director of Research & Innovation Office, at [email protected].
