Shallow Recurrent Decoder for Aero-Optical Wavefront Sensing

Abstract

We demonstrate SHRED, a SHallow REcurrent Decoder, to augment aero-optical wavefront sensors. The neural network achieves high-fidelity state reconstruction and forecasting using sparse, robust sensor placements and is useful for adaptive optic predictive control.

Full text

Acknowledgements: This work was supported by the Air Force
Office of Scientific Research (FA9550- 19-1-0386) and the US
National Science Foundation AI Institute for Dynamical Systems
(dynamicsai.org), grant 2112085. In addition, the authors
acknowledge support by the Air Force Research Laboratory
Directed Energy Directorate by way of access to experimental data
and applications from their facilities.
Approved for public release; distribution is unlimited. Public Affairs release approval #AFRL-2023-4809
References
[1] Sahba, S., Sashidhar, D., Wilcox, C. C., McDaniel, A., Brunton, S. L., & Kutz, J. N.
(2022). Dynamic mode decomposition for aero-optic wavefront characterization.
Optical Engineering, 61(1), 013105-013105.
[2] Williams, J. P., Zahn, O., & Kutz, J. N. (2023). Sensing with shallow recurrent
decoder networks. arXiv preprint arXiv:2301.12011.
[3] Gordeyev, S., Jumper, E. J., & Whiteley, M. R. (2023). Aero-Optical Effects: Physics,
Analysis and Mitigation. John Wiley & Sons.
Results & Outlook
Sub-millisecond aero-optical wavefront forecasting in highly
turbulent conditions is possible with SHRED. Aero-optical
aberrations are sensitive to beam direction and airborne optical
platform geometry and elevation, yet SHRED provides a lightweight
framework for quick reconfiguration to serve as a low latency
predictor for adaptive optics control. Additionally, due to SHRED's
sparse sensor placement requirement, there is inherent possibility
for data compression when recording and reconstructing high-
fidelity wavefronts.
Figure 4. Truth, reconstruction, and forecasted snapshots from SHRED on a AAOL-T test set. The
inputs are 12 radially dispersered sensor recordings, each with a 60 snapshot memory (~2 ms).
Reshaped wavefront examples at t=200 are shown. The forecasted time series for the 12 sensors
are shown together and again separately for a shorter, relevant timeframe for adaptive optics.
Figure 3. SHRED network for aero-optical flow reconstruction from sensor point measurements.
Sparse sensor measurements are provided to an LSTM network then fed into a shallow decoder
to devleop two models: one for sensor forecasting and another for full state reconstruction. In
tandem, we achieve high-fidelity forecasting suitable for adaptive optics in turbulent scenarios.
SHRED for Wavefront Reconstruction and Forecasting via Sparse Sensor Placements
Aero-Optics with In-flight Data
In airborne optical environments, aero-optical effects such as
turbulent boundary layers cause rapid index of refraction
fluctuations that induce aberrations in transmitted wavefronts.
The fast-scale physics in the turbulent flows lead to an extremely
low latency requirement for control. Computationally efficient
forecasting is critical in free space optical systems [1]. Thus we
propose SHRED [2], given its ability to rapidly forecast turbulent
flow fields for short-term predictive control.
Figure 2. (left) Schematic of a Shack-Hartmann wavefront sensor, which are used in the AAOL-T
and reconstruct incident wavefronts from "local tilts" generated by beamlet spot displacements
onto a sensor array. (right) SHRED is a candidate predictive controller for an adaptive-optics
loop, which requirs low latency forecasts to correct beam aberrations via a deformable mirror.
Figure 1. We study in-flight wavefront transmission
data from the Airborne Aero-Optics Laboratory
Transonic or AAOL-T [3].
Shervin Sahba1, Christopher C. Wilcox2, Austin McDaniel2, Benjamin D. Shaffer2, J. Nathan Kutz3
Shallow Recurrent Decoder for Aero-Optical
Wavefront Sensing and Forecasting
1Department of Physics, University of Washington, Seattle, WA
2US Air Force Research Laboratory, Kirtland Air Force Base, NM
3Department of Applied Mathematics, University of Washington, Seattle, WA
Abstract
SHRED is a SHallow REcurrent Decoder neural
network. In aero-optics, where beam transmission
meets turbulent aerodynamic flow, SHRED, provided
the time histories of sparsely placed sensors, can
reconstruct and forecast aero-optical wavefronts.