Dynamic mode decomposition based predictive model performance on supersonic and transonic aero-optical wavefront measurements

Abstract

Air density variations around an airborne directed energy system distort a beam’s wavefront, resulting in degraded performance after propagation into the far field. Adaptive optics (AO) can be used to correct for these rapidly evolving aero-optical aberrations; however, in some conditions, the inherent latency between measurement and correction in state-of-the-art AO systems results in significantly reduced performance. Predictive AO control methods utilize future state predictions to compensate for rapidly evolving distortions and are promising techniques for mitigating this limitation. This paper demonstrates an application of the dynamic mode decomposition (DMD) method on turbulent boundary layer wavefront data from supersonic and transonic wind tunnel flow from the Air Force Research Laboratory’s Aero-Effects Laboratory. DMD is a lightweight algorithm used to isolate spatiotemporal patterns in a dataset into physically meaningful modes with associated dynamics, which were used to predict future states from a given wavefront. This method showed notable improvements in simulated wavefront correction, providing a reduction of residual wavefront distortion, measured as root mean square over the aperture, by up to 25.4% over a simulated latency model, which could accordingly result in higher laser system performance.

Neural network forecasting of transonic turbulent flow for adaptive optics control

“…Artificial Neural Networks are a somewhat natural choice as the nonlinear embedding allows a relatively simple model to achieve the complicated task of first determining the flow dynamics for different data conditions, then being able to apply appropriate dynamics on input data to produce a relevant prediction. Linear models such as Vector Autoregression and DMD have demonstrated strong prediction capability [11][12][13][14][15][16][17][18]. However, these methods fundamentally apply a single, constant time evolution on the input wavefront to produce the output, with no built in mechanism for embedding or applying dynamics corresponding to different data conditions, and therefore are less suitable for a generalizable model.…”