Machine Learning Reconstruction of Compressible Turbulent Bluff-Body Flow

The accurate reconstruction of high-dimensional spatiotemporal flow fields in compressible turbulent wakes is relevant for developing hypersonic vehicles. However, the inherent complex multiscale dynamics of compressible wakes and the practical constraints of high-speed environments often limit in-situ measurements to sparse, point-based sensors. As such, standard methods often entail a sparse number of point-based sensors that can only determine the flow field's temporal nature but lack spatial content. This work addresses these challenges by leveraging a physics-constrained Shallow Recurrent Decoder (SHRED) neural architecture, to reconstruct complete flow fields from limited sensor data. Specifically, this research utilizes experimental bluff-body wake data spanning Reynolds numbers from 5,000 to 1,000,000 for training and ground truth validation. Results show that as few as three probes are sufficient to produce high-fidelity reconstructions of the flow field. This approach demonstrates robust, accurate flow reconstruction from sparse measurements, advancing real-time monitoring, control, and the development of digital twins for complex, high-speed aerodynamic systems.