RECORD
Transformer neural networks for closed-loop adaptive optics using non-modulated pyramid wavefront sensors
Instrumentation and Methods for Astrophysics · Submitted 9 May 2024
abstract compiled · 2026-09-23 · no language-model review
The authors' main claim In conclusion, we demonstrate that a TNN is the most suitable architecture to extend the dynamic range without sacrificing sensitivity for a non-modulated PyWFS.
The compiled record
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
Open the compiled record
The abstract, sentence by sentence
The Pyramid Wavefront Sensor (PyWFS) is highly nonlinear and requires the use of beam modulation to successfully close an AO loop under varying atmospheric turbulence conditions, at the expense of a loss in sensitivity. In this work we train, analyse, and compare the use of deep neural networks (NNs) as non-linear estimators for the non-modulated PyWFS, identifying the most suitable NN architecture for reliable closed-loop AO. We develop a novel training strategy for NNs that seeks to accommodate for changes in residual statistics between open and closed-loop, plus the addition of noise for robustness purposes. Through simulations, we test and compare several deep NNs, from classical to new convolutional neural networks (CNNs), plus a state-of-the-art transformer neural network (TNN, Global Context Visual Transformer, GCViT), first in open-loop and then in closed-loop. Using open-loop simulated data, we observe that a TNN (GCViT) largely surpasses any CNN in estimation accuracy in a wide range of turbulence conditions. Also, the TNN performs better in simulated closed-loop than CNNs, avoiding estimation issues at the pupil borders. When closing the loop at strong turbulence and low noise, the TNN using non-modulated PyWFS data is able to close the loop similar to a PyWFS with $12\lambda/D$ of modulation. When raising the noise only the TNN is able to close the loop, while the standard linear reconstructor fails, even with modulation. Using the GCViT, we close a real AO loop in the optical bench achieving a Strehl ratio between 0.28 and 0.77 for turbulence conditions ranging from 6cm to 20cm, respectively. In conclusion, we demonstrate that a TNN is the most suitable architecture to extend the dynamic range without sacrificing sensitivity for a non-modulated PyWFS. It opens the path for using non-modulated Pyramid WFSs under an unprecedented range of atmospheric and noise conditions.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | The Pyramid Wavefront Sensor (PyWFS) is highly nonlinear and requires the use of beam modulation to successfully close an AO loop under varying atmospheric turbulence conditions, at the expense of a loss in sensitivity. | background 1.00 | 0.07 | 0.02 | asserted 1.00 |
| s1 | In this work we train, analyse, and compare the use of deep neural networks (NNs) as non-linear estimators for the non-modulated PyWFS, identifying the most suitable NN architecture for reliable closed-loop AO. | contribution 0.99 | 0.95 | 0.04 | asserted 0.92 |
| s2 | We develop a novel training strategy for NNs that seeks to accommodate for changes in residual statistics between open and closed-loop, plus the addition of noise for robustness purposes. | method 0.80 | 0.95 | 0.02 | asserted 0.98 |
| s3 | Through simulations, we test and compare several deep NNs, from classical to new convolutional neural networks (CNNs), plus a state-of-the-art transformer neural network (TNN, Global Context Visual Transformer, GCViT), first in open-loop and then in closed-loop. | method 0.94 | 0.94 | 0.03 | asserted 0.90 |
| s4 | Using open-loop simulated data, we observe that a TNN (GCViT) largely surpasses any CNN in estimation accuracy in a wide range of turbulence conditions. | result 1.00 | 0.96 | 0.10 | asserted 1.00 |
| s5 | Also, the TNN performs better in simulated closed-loop than CNNs, avoiding estimation issues at the pupil borders. | result 1.00 | 0.95 | 0.06 | asserted 1.00 |
| s6 | When closing the loop at strong turbulence and low noise, the TNN using non-modulated PyWFS data is able to close the loop similar to a PyWFS with $12\lambda/D$ of modulation. | result 1.00 | 0.95 | 0.10 | asserted 0.99 |
| s7 | When raising the noise only the TNN is able to close the loop, while the standard linear reconstructor fails, even with modulation. | result 1.00 | 0.93 | 0.06 | asserted 1.00 |
| s8 | Using the GCViT, we close a real AO loop in the optical bench achieving a Strehl ratio between 0.28 and 0.77 for turbulence conditions ranging from 6cm to 20cm, respectively. | result 1.00 | 0.97 | 0.07 | asserted 1.00 |
| s9 | In conclusion, we demonstrate that a TNN is the most suitable architecture to extend the dynamic range without sacrificing sensitivity for a non-modulated PyWFS. | interpretation 0.78 | 0.90 | 0.96 | asserted 1.00 |
| s10 | It opens the path for using non-modulated Pyramid WFSs under an unprecedented range of atmospheric and noise conditions. | interpretation 0.93 | 0.43 | 0.15 | asserted 0.66 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
The paper
Original abstract
The Pyramid Wavefront Sensor (PyWFS) is highly nonlinear and requires the use of beam modulation to successfully close an AO loop under varying atmospheric turbulence conditions, at the expense of a loss in sensitivity. In this work we train, analyse, and compare the use of deep neural networks (NNs) as non-linear estimators for the non-modulated PyWFS, identifying the most suitable NN architecture for reliable closed-loop AO. We develop a novel training strategy for NNs that seeks to accommodate for changes in residual statistics between open and closed-loop, plus the addition of noise for robustness purposes. Through simulations, we test and compare several deep NNs, from classical to new convolutional neural networks (CNNs), plus a state-of-the-art transformer neural network (TNN, Global Context Visual Transformer, GCViT), first in open-loop and then in closed-loop. Using open-loop simulated data, we observe that a TNN (GCViT) largely surpasses any CNN in estimation accuracy in a wide range of turbulence conditions. Also, the TNN performs better in simulated closed-loop than CNNs, avoiding estimation issues at the pupil borders. When closing the loop at strong turbulence and low noise, the TNN using non-modulated PyWFS data is able to close the loop similar to a PyWFS with $12\lambda/D$ of modulation. When raising the noise only the TNN is able to close the loop, while the standard linear reconstructor fails, even with modulation. Using the GCViT, we close a real AO loop in the optical bench achieving a Strehl ratio between 0.28 and 0.77 for turbulence conditions ranging from 6cm to 20cm, respectively. In conclusion, we demonstrate that a TNN is the most suitable architecture to extend the dynamic range without sacrificing sensitivity for a non-modulated PyWFS. It opens the path for using non-modulated Pyramid WFSs under an unprecedented range of atmospheric and noise conditions.
Human review
How the literature uses it
2 arXiv papers cite this in 5 sentences
Papers citing it · 3 sentences without a confident label
Cites it as background 2 papers
-
Previous studies have used NN to reconstruct the signal from a PWFS [ref] , with the first on-sky demonstration by [ref] .
Closing the loop on-sky with a vector-Zernike wavefront sensor using a convolutional neural network as phase reconstructor · Introduction · 0.82
-
Many approaches have been proposed to overcome the nonlinearity of both the uPWFS and the modulated PWFS, ranging from first-order optical-gain compensation [ref] to model-based nonlinear reconstructors [ref] and data-driven reconstructors based on neural networks (NNs) [ref] .
No need to modulate: On-sky results of a Neural Network enhanced pyramid wavefront sensor and prospects for the ELTs · INTRODUCTION · 0.88
What it cites · 3 sentences without a confident label
Uses the method or tool of 2 papers
-
Xception [ref] , WFNet [ref] , ConvNext [ref] , and GCVit [ref] , are implemented in PyTorch and adapted to perform regression at the last layer to provide simultaneous estimations for the first $209$ Zernike modes (without piston).
A ConvNet for the 2020s · Neural network training · 0.95
-
Xception [ref] , WFNet [ref] , ConvNext [ref] , and GCVit [ref] , are implemented in PyTorch and adapted to perform regression at the last layer to provide simultaneous estimations for the first $209$ Zernike modes (without piston).
Global Context Vision Transformers · Neural network training · 0.95
Cites as background 3 papers
-
Nevertheless, there is a new generation of CNNs–ConvNeXt and ConvNeXt v2 [ref] –that are delivering impressive performance in classification tasks.
A ConvNet for the 2020s · Introduction · 0.90
-
Nevertheless, there is a new generation of CNNs–ConvNeXt and ConvNeXt v2 [ref] –that are delivering impressive performance in classification tasks.
ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders · Introduction · 0.90
-
Most of the deep learning WFSs have been adapting conventional deep neural nets originally developed for computer vision applications, such as Xception [ref] , VGG–Net [ref] , or ResNet [ref] , while in [ref] the authors crafted an original deep neural net (WFNET) for image-based WFSs.
1409.15564 · Introduction · 0.97
Discussion
Know something about this paper? Leave a comment, or highlight any passage above to comment on it.