PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...

Page created by Terrence Dominguez
 
CONTINUE READING
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
PCS R&D – predictive control
for eXtreme Adaptive Optics
 M. Kasper
 J. Nousiainen, N. Cerpa Urra, M. Bonse,
P. Pathak, S. Leveratto, P. Bristow, T. Helin,
 C. Kulcsar, H.F. Raynaud, S. Quanz
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
ELT – Planetary Camera and Spectrograph
 (PCS)
Science: Characterize nearby Earth-like Exoplanets, find biomarkers
Concept: eXtreme Adaptive Optics (XAO) + high-resolution spectroscopy
Time-scale: ongoing R&D, Project Start ~2025, 1st light ~2035

 PR eso1629de
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
Exoearths are VERY faint
 and difficult to observe
 .. Illustration, Kate Follette

 .
 .
 0.00003°
 5ft

Like a firefly from 3000 km away
…and sitting next to a lighthouse beam
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
Ground-based telescopes look through
 the atmosphere

EiroForum Big Data, Oct 2020
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
Adaptive optics (AO) corrects
 atmospheric turbulence

 Delay between wavefront and
 correction by DM: > 2 timesteps

EiroForum Big Data, Oct 2020
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
Deformable Mirrors

 Three technologies:
 • Voice-coil DSMs
 • Stacked Piezo mirrors
 • Electro-static MEMS

 Aperture: 10 mm – 1 m
 Actuator pitch: 0.3 – 30 mm
 Number of actuators: < 4000

EiroForum Big Data, Oct 2020
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
Wave-front Sensors

 SHS: high-linearity, LGS
 PWS: high-sensitivity, XAO
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
XAO error budget dominated by time delay

 Time delay

 Plot by C. Verinaud
 −5Τ3
 ∝ ℎ ≈ 0 × 3 
EiroForum Big Data, Oct 2020
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
AO delay on sky (VLT-SPHERE)

 High wind
 V = 22 ± 4m/s

 Diagonal elongation in wind direction
 (faster decorrelation of turbulence)
Low wind
V = 3 m/s ± 1 m/s
EiroForum Big Data, Oct 2020
PCS R&D - predictive control for eXtreme Adaptive Optics - M. Kasper J. Nousiainen, N. Cerpa Urra, M. Bonse, P. Pathak, S. Leveratto, P. Bristow ...
How to reduce time delay? Option 1
 Run a faster 2nd AO stage (PhD Nelly Cerpa Urra)

 2nd stage boosts contrast, increased framerate
EiroForum Big Data, Oct 2020 increases noise and reduces sensitivity
Option 2: Predictive control
Simulated turbulence Multi Layer, frozen flow Measurement SPHERE
Single Layer, frozen flow Contra-directional wind On-sky

Very high spatio-temporal Low spatial correlation. Boiling, non-stationary
correlation. More difficult to predict More difficult to predict
Easy to predict
 Record lots of sky turbulence (AO telemetry)
 VLT/SPHERE (~4000 DoF, 1.4 kHz): ~10 GB/min
 ELT: ~250 GB/min
Temporal correlation not affected by
 multiple layers
Use Reinforcement Learning for AO
Model AO system dynamics as a Markov decision process (MDP)
Use Neural Network to learn MDP transition probabilities (AO dynamics),
consider 4 previous and 2 future time steps
Use a planning algorithm to determine next DM action, which maximizes
reward (minimizes WFS residuals)

 Nousiainen et al.
 In preparation
RL predicts wavefront evolution

 No shift between RL an ‘no delay’
 Integrator is lagging behind
EiroForum Big Data, Oct 2020
RL significantly improves contrast
 over a wide range of noise levels
No noise = infinitely bright AO guide star

EiroForum Big Data, Oct 2020
RL significantly improves contrast over
 a wide range of noise levels
Noisy case = relativey faint AO guide star

EiroForum Big Data, Oct 2020
RL may have additional benefits
WFS non-linearities (e.g. non-modulated PWS) Mis-registration DM vs WFS
 Temporal variation during observation
 s

 f

 Non-linear method, interacting with system and adjusting in real-time
 EiroForum Big Data, Oct 2020
RL copes with mis-registration

 Lateral shift between DM
 and WFS has no effect

EiroForum Big Data, Oct 2020
RL converges quickly

 RL outperforms Integrator after a few seconds
EiroForum Big Data, Oct 2020
Summary and conclusion

 PCS will be the ELT’s Exoearth characterization instrument
 Fast XAO is an enabling technology for this science case
 Predictive control with Reinforcement Learning offers a path
 to a highly performant turn-key system
 Rapid progress in algorithms and HW promise a feasible
 implementation in PCS in about a decade (2030+)
 Next steps: R&D on simulations with realistic turbulence, on a
 bench setup (2021-22) and on sky (~2023-24)

EiroForum Big Data, Oct 2020
You can also read