Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor

summary

The gist

Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor is shown to extend the dynamic range of a sensitive wavefront sensor far…

In short

The study used a Convolutional Neural Network (CNN) as a phase reconstructor to close the loop on-sky with a vector-Zernike wavefront sensor (v-ZWFS). This technique successfully extended the sensor's dynamic range far beyond linear limits, allowing stable operation in conditions where traditional linear methods failed. The CNN achieved higher Strehl ratios than linear reconstructors.

Key concepts

Vector-ZWFS (v-ZWFS)
This is a high dynamic range version of the Zernike wavefront sensor that uses two Zernike masks with different phase shifts. This design allows the sensor to capture a wider range of aberrations, effectively doubling its linear operating range compared to standard sensors.
Convolutional Neural Network (CNN)
A type of artificial intelligence used here as a non-linear phase reconstructor. It takes two images representing the v-ZWFS input and learns the complex relationship between these inputs and the desired output phase, allowing it to perform reconstructions beyond what simple linear math can achieve.
Log-Variance Loss Function
This is a specific mathematical loss function used during CNN training. It prioritizes minimizing errors in regions where small residuals are physically important (low residual variance) while compressing the error signal in high-residual regimes, helping the network learn to accurately reconstruct both large and small phase errors.
On-sky Demonstration
This refers to testing the trained CNN reconstructor directly on a real instrument observing conditions. The results showed that the CNN consistently maintained higher Strehl ratios compared to linear methods when both were tested under similar atmospheric turbulence conditions, proving its practical viability.

Terminology used across episodes

This episode discusses

The paper

Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor · Read on arXiv

F. Oyarzún, M. Motte, C. Heritier, B. González, R. Muñoz, V. Chambouleyron, M.A.¡lagao, Á Striffling M., Ο Pasinetti E., Ο Vinerskas P., T. Trouve-Peloux P., F. Champagnat F., T. Fusco T., B. Neichel B.

Aix Marseille University · CNRS · CNES · LAM · DOTA · “ Paris Saclay Department of Technology and Research Applications, ONERA · – School of Electrical Engineering, Pontificia Universidad Catolica de Valparaiso

Context: The new giant segmented mirror telescopes will use adaptive optics to reach the fundamental limits in resolving power. To accomplish this, new wavefront sensors (WFS) have been designed to fulfill the requirements, but they may require non-linear reconstruction techniques to operate given the nature of the signal of the WFSs Aims: In this article we show that it is possible to use non-linear reconstructors to extend the dynamic range of one of the most sensitive wavefront sensors far beyond the designed limits: the Zernike wavefront sensor (ZWFS). Methods: We trained a convolutional neural network (CNN) completely in simulation to perform the phase reconstruction of a vector-ZWFS (v-ZWFS), a higher dynamic-range variant of the ZWFS. Contrary to the linear method, the CNN uses the information across the full frame to reconstruct the phase at each point, enabling it to resolve the ambiguities introduced by the periodic response of the ZWFS and thereby extend its effective capture range. We developed a two-step training strategy that ensured closed-loop stability and used a physically informed loss function to maximize the performance of the CNN. Results: We successfully closed the loop on-sky with the v-ZWFS using the CNN in observing conditions that the linear reconstructor could not converge to a stable flat wavefront. In cases where both reconstruction methods were working, the CNN outperformed the linear method in almost all cases, and in favorable seeing conditions we were even able to close the loop with the ZWFS acting as a first stage WFS, highlighting the extended dynamic range brought by the use of a non-linear wavefront reconstructor. Conclusions: We conclude that the use of non-linear wavefront reconstructor can extend the use cases of adaptive optics systems, especially when the WFS used shows highly non-linear behaviors.

DOI: 10.1051/0004-6361/202661297

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor".

Jocelyn: Here's the gist of the paper. Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor is shown to extend the dynamic range of a sensitive wavefront sensor far beyond designed limits through non-linear reconstruction techniques The CNN successfully closed the loop on-sky with the v-ZWFS using observing conditions that the linear reconstructor could not converge to a stable flat wavefront Context and Motivation The paper addresses the need for non-linear reconstruction techniques to operate highly sensitive wavefront sensors like the Zernike wavefront sensor (ZWFS) given their inherent signal nature Direct imaging of exoplanets will be a main objective for new giant telescopes, but ground-based telescopes suffer resolution loss due to atmospheric turbulence, degrading effective resolution to θseeing = λ/r0 Fourier Filtering Wavefront Sensors (FFWFSs), such as the Zernike (ZWFS), are attractive for their high sensitivity, enabling observations of fainter guide stars and increasing sky coverage While the relation between input phase and intensity can be linearized for small phase amplitudes, the nonlinear response at larger aberrations limits standard reconstruction techniques Vector-ZWFS Modeling and Data Generation The study focuses on the vector-ZWFS (v-ZWFS), a higher dynamic-range variant of the ZWFS, which is composed of two Zernike masks with different phase shifts The signal from each mask has different proportions of sine and cosine, which can double the linear range of the sensor To generate training data, a power spectral density (PSD) for atmospheric wavefronts is constructed by combining the PSD of the atmosphere, modeled as Kolmogorov or Von-Karman distribution, with PSDs from the deformable mirror and temporal errors The dataset contains wavefronts with various conditions including r0 between 8 - 20 cm, wind speeds ranging from-15 to 15 m/s, and residuals representing open-loop, closed-loop, and intermediate statistics CNN Architecture and Training Strategy A Convolutional Neural Network (CNN) was used for phase reconstruction, taking two channels as input corresponding to the two pupil images of the vZWFS The CNN architecture consisted of six convolutional layers with 16 to 512 filters, applying leaky ReLU activation and 2 × 2 max-pooling after each layer To mitigate the issue where the network prioritized low order modes, curriculum training was used by shifting training to power laws that better balanced the energy distribution across the modes The loss function proposed is L = 1/n sum Xn i=1 log var(ϕres), which compresses the high-residual-phase regime while expanding the low-residual-phase regime Performance and Conclusion The simulation results demonstrated that the CNN reconstructor extends the usable capture range well beyond that of the linear method in both open-loop and closed-loop operation In closed-loop operation, the CNN kept residual variance below 10% of the input for all tested variances greater than 1 rad squared, whereas the linear reconstructor approached the unity line above 1 rad squared On-sky demonstrations showed that the CNN consistently achieved higher Strehl ratios when both reconstructors were operated under comparable conditions The work concludes that a simulation-trained nonlinear reconstructor can be deployed directly on-sky without retraining on real data, which is critical for future instruments Key Contributions The key contributions of this work include: 1.

Vera: An accurate instrument model accounting for DM misregistration 2.

Jocelyn: A log-variance loss function that encodes the physical importance of small residuals 3.

Vera: A two-step training strategy that ensures closed-loop stability These findings suggest that nonlinear reconstruction is a qualitative shift in the use cases available to highly sensitive but inherently nonlinear wavefront sensors such as the ZWFS The CNN consistently outperformed the linear method when both methods converged, enabling stable closed-loop operation in conditions where the linear approach failed entirely The paper emphasizes that fidelity of the instrument model is more critical to on-sky success than the specific CNN architecture The performance limits of AO systems can be significantly extended through nonlinear reconstruction methods Acknowledgements This work benefited from the support the French National Research Agency (ANR) with the Programme Investissement Avenir F-CELT (ANR-21-ESRE-0008), PEPR ORIGINS (XAO-WFS ANR-22-EXOR-00062, and COMPACT SPECTROGRAPHS ANR-22-EXOR-0006), the ANR DGAAID ASTRID program (ANR-25-ASTR-0015), the Action Spécifique Haute Résolution Angulaire (ASHRA) of CNRS/INSU co-funded by CNES, the french government under the France 2030 investment plan (cassiopée project) and the Initiative d’Excellence d’Aix-Marseille Université A MIDEX Authors also would like to thank Kent Wallace and Tobias Wegner for their support on the vector-ZWFS References Alagao, A., Oyarzun, F., & Chambouleyron, V.

Paper summary: Vera: The main idea here is that you can use these non-linear reconstruction techniques to extend the usable dynamic range of your sensitive wavefront sensor far beyond what standard linear methods can handle.

Jocelyn: Specifically, they focus on the Zernike wavefront sensor, which has this signal nature that makes it tricky at larger phase amplitudes.

Vera: The paper claims that a convolutional neural network can successfully close the loop on-sky with this vector-Zernike wavefront sensor even under observing conditions where a linear reconstructor just couldn't settle on a stable flat wavefront.

Subrahmanyan: It’s important that they are using non-linear reconstruction because the linear response at large phase amplitudes causes ambiguities, like getting the wrong sign for the phase estimate, which can drive things into instability (Chambouleyron et al. 2021b; Striffling et al <ref:2608.07379#pg2>. two thousand twenty-five) <ref:2608.07379#pg1>.

Jocelyn: They are using a vector-Zernike wavefront sensor, which is actually a higher dynamic range variant of the standard Zernike sensor, consisting of two Zernike masks with different phase shifts, and this doubles the linear range of the sensor.

Vera: To train their convolutional neural network for this task, they built a dataset that combined atmospheric wavefronts modeled by things like Kolmogorov or Von Karman distributions with errors from the deformable mirror and temporal errors.

Subrahmanyan: They also focused on modeling the deformable mirror, introducing misregistration between it and the wavefront sensor to simulate real-world instrumental challenges, which is something they say leads to model mismatch and possible instabilities during closed loop operation (Heritier et al. two thousand eighteen) <ref:2608.07379#pg1>.

Vera: The network itself has six convolutional layers with filters ranging from sixteen up to five hundred twelve, using leaky ReLU activation and two by two max-pooling after each layer.

Jocelyn: And they used a specific loss function, which is this log variance loss function: L equals one over n sum X sub n i equals one log var of the residual phase, which compresses the high residual phase regime while expanding the low residual phase regime <ref:2608.07379#pg1>.

Subrahmanyan: It’s interesting that they used curriculum training to shift their training to power laws that better balanced how the energy is distributed across all those different modes.

Vera: The results show that this CNN reconstructor extends the usable capture range well beyond what the linear method achieves in both open-loop and closed-loop operation.

Jocelyn: In closed-loop operation, they kept the residual variance below ten percent of the input for all tested variances greater than one rad squared, while a linear reconstructor approached unity above that same value.

Subrahmanyan: This suggests that this approach is robust enough to handle those conditions where traditional linear methods fail entirely, which is what you need for real adaptive optics systems.

Conclusion: Vera: So, looking at the title, "Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor", it really tells you that they managed to solve a specific problem: getting those sensitive sensors to work reliably outside of ideal conditions by using this AI reconstruction method.

Jocelyn: They are showing that this CNN reconstructor can be deployed directly on-sky without needing further retraining on real data, which is pretty critical for instruments that are operating in real time.

Subrahmanyan: The big implication here is that nonlinear reconstruction isn't just some academic exercise; it shows a way to significantly extend the performance limits of adaptive optics systems by using neural networks trained in simulation.

Vera: It’s about moving the capability of these highly sensitive but inherently nonlinear sensors from being limited by their physics to being limited by how well you model them and train your network.

Jocelyn: The work emphasizes that the fidelity of the instrument model, specifically accounting for things like deformable mirror misregistration, seems more critical to making this on-sky success than the specific architecture of the CNN itself.

Subrahmanyan: They suggest that these nonlinear reconstruction methods are a qualitative shift in what's possible when you use wavefront sensors like Zernike sensors for next-generation telescopes.

Vera: It means that we can potentially push the performance limits of adaptive optics systems much further than before by employing these kinds of non-linear reconstruction techniques.

Jocelyn: This allows us to operate in conditions where the linear approach simply cannot converge to a stable flat wavefront, which is exactly what you need for high-resolution imaging.

Subrahmanyan: The challenge for future work will be making sure that this simulation-trained reconstructor performs reliably when it’s actually pointed at a real sky, not just simulated data.

Vera: So, to wrap up on this paper, they've shown a way to use deep learning to handle the non-linear response of a vector-Zernike wavefront sensor and close the loop on-sky successfully.

Jocelyn: It’s a demonstration that these methods can extend the usable capture range significantly compared to purely linear approaches.

Subrahmanyan: The paper sets the stage for more robust AO systems where we can use these types of reconstructors in real instruments.

More episodes

← Home