A deep learning-based aberration compensation multi-deformation mirror voltage control method

Abstract

The application discloses a kind of based on deep learning's aberration compensation multi-deformation mirror voltage control method, belong to strong turbulent aberration correction technical field.The method first simulates multi-layer atmospheric turbulence, constructs simulation dataset including multi-layer phase screen, pupil plane intensity image and Shack-Hartmann image;Second, build deep learning network based on physical information, with pupil plane intensity image and Shack-Hartmann image as input, directly output the control voltage of multi-layer deformable mirror;According to the physical law of light wave transmission, the combined loss function composed of pupil plane scintillation index and far field Strehl ratio is designed to train the network, to solve the problem of multi-deformable mirror correction aberration coupling and restored aberration uncertainty.The application can calculate the optimal voltage in millisecond level, avoid the error accumulation of traditional method, significantly improve the imaging quality under strong turbulent, and has important significance for high-resolution imaging observation.

Description

Technical Field

[0001] This invention belongs to the field of strong turbulent aberration correction technology, specifically relating to a voltage control method for aberration compensation of deformable mirrors based on deep learning. Background Technology

[0002] High-resolution imaging of celestial targets by ground-based optical telescopes is severely limited by atmospheric turbulence. Adaptive Optics (AO) technology effectively improves the imaging resolution of telescopes by detecting and correcting wavefront distortions caused by atmospheric turbulence in real time, and has become an indispensable component of large ground-based telescopes. A typical AO system includes a wavefront sensor, a wavefront controller, and a wavefront corrector, among which the Shak-Hartmann wavefront sensor is widely used due to its mature structure and fast detection speed.

[0003] Currently, AO (Analog-Optical Front) systems can operate stably under weak turbulence. However, under medium to high-intensity turbulence, atmospheric scintillation causes severe unevenness in the pupil light intensity distribution, resulting in intensity flicker and even phase singularities. This directly leads to excessively low or missing spot energy within some sub-apertures of the Shaker-Hartmann wavefront sensor, causing the centroid detection algorithm to fail and resulting in a sharp decline in wavefront reconstruction accuracy, rendering the AO system unstable. Furthermore, traditional wavefront reconstruction methods (such as mode methods and region methods) are sensitive to system assembly errors and non-common-path aberrations, further limiting their detection performance in complex environments, and they also cannot solve the problem of performance degradation of AO systems under strong turbulence.

[0004] Existing research on aberration correction under strong turbulence mainly falls into two categories. One category employs iterative algorithms such as the GS algorithm or SPGD algorithm to approximate the target wavefront through multiple iterations. However, limited by iteration accuracy and control bandwidth, iterative algorithms cannot be effective in practical applications. The other category introduces deep learning technology to establish a nonlinear mapping between the input image and the target wavefront through a data-driven approach, which improves accuracy compared to traditional methods. However, existing methods still have two key problems: First, under strong turbulence, the pupil surface light intensity flickers and the phase is discontinuous, with phase singularities. Even if the wavefront can be recovered, the deformable mirror cannot fit this discontinuous wavefront. Second, the pupil surface light intensity distribution is uneven, and even if the phase is corrected, the far-field Strell ratio still cannot reach the ideal value.

[0005] More importantly, current AO (Anatomical Reduction) techniques under strong turbulence are still limited to the restoration of the pupil plane phase, neglecting the correction of pupil plane intensity scintillation itself, and cannot fundamentally address the phase singularity problem, thus failing to fundamentally solve the problems faced by AO systems under strong turbulence. Furthermore, existing techniques suffer from restoration errors during wavefront restoration, and the application of the wavefront to the deformable mirror introduces surface shape fitting errors; these two errors lead to even greater system errors. Therefore, a new method is urgently needed that can directly output the deformable mirror voltage, avoid error accumulation, and simultaneously correct intensity scintillation and wavefront distortion, fundamentally solving the challenges faced by AO systems under strong turbulence. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a deep learning-based method for voltage control of multiple deformable mirrors with aberration compensation. Multiple deformable mirrors are conjugated to different heights to correct atmospheric turbulence. A deep learning network based on physical information is constructed, with pupil plane light intensity images and Shack-Hartmann images as inputs. The network is trained using a combined loss function consisting of the pupil plane scintillation index and the far-field Strell ratio, which directly outputs the voltages of multiple deformable mirrors, thereby achieving millisecond-level, error-free direct aberration compensation under strong turbulence.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A voltage control method for aberration compensation of a deformable mirror based on deep learning, the method comprising the following steps:

[0009] A voltage control method for aberration compensation of a deformable mirror based on deep learning, the method is implemented through the following steps:

[0010] Step 1: Use a multi-layer turbulent phase screen to simulate real atmospheric turbulence, and combine it with the angular spectrum diffraction transmission method to establish a layered transmission model of light waves in atmospheric turbulence;

[0011] Step 2: Based on the aforementioned atmospheric turbulence layered transmission model, construct a simulation dataset. Each set of simulation data includes a layered turbulence phase screen, a pupil light intensity image obtained after transmission, and a Shaker-Hartmann image.

[0012] Step 3: Embed the physical model of light wave atmospheric turbulence after layered transmission and correction by multiple deformable mirrors into the deep learning network, build a deep learning network based on physical information, and construct a combined loss function composed of pupil light intensity scintillation index and far-field Strell ratio;

[0013] Step 4: Using the pupil light intensity image and the Shaker-Hartmann image from the simulation dataset as network inputs, train the constructed physical information-based deep learning network. During the training process, substitute the deformable mirror voltages predicted by the network into the physical model to calculate the combined loss function, update the network parameters, and finally output the voltages of multiple deformable mirrors.

[0014] Furthermore, the parameters of the multi-layer turbulent phase screen mentioned in step 1 are obtained based on the atmospheric turbulence stratification theory.

[0015] Furthermore, the physical model embedded in the deep learning network in step 3 is used to describe the process of wavefront distortion correction after the beam passes through the turbulent phase screen for layered diffraction transmission and is then loaded by multiple deformable mirrors, in which pupil light intensity and far-field images are obtained; wherein the number of turbulent phase screens and the number of deformable mirrors are set according to the layered correction requirements.

[0016] Furthermore, the combined loss function in step 3 is composed of a weighted sum of a pupil intensity scintillation index term and a far-field Strell ratio term; wherein the far-field Strell ratio is used to evaluate the quality of the far-field spot after correction by the deformable mirror, and the pupil intensity scintillation index is used to evaluate the uniformity of pupil intensity after correction by the deformable mirror, and the sum of the weighting coefficients of the two terms is one.

[0017] Furthermore, the combined loss function is used to correct high-level turbulence using a high-level deformable mirror and correct surface turbulence using a surface deformable mirror, thereby avoiding uncertainties in the coupling of multi-deformable mirror aberration correction and the restoration of aberrations.

[0018] Furthermore, turbulent aberrations at different altitudes collectively affect the pupil light intensity distribution in a nonlinear manner, and the Shaker-Hartmann image is used to reflect the turbulent aberrations of the entire atmospheric layer.

[0019] Furthermore, the training process in step 4 is as follows: after normalizing the pupil light intensity image and the Shaker-Hartmann image, the network input is used; the deformable mirror voltage predicted by the network is substituted into the physical model; the loss of each batch of data is calculated through the combined loss function; the gradient of the loss with respect to each parameter in the network is calculated and the parameters are updated until the loss converges.

[0020] Furthermore, the trained physical information-based deep learning network is used to output the voltages of multiple deformable mirrors in milliseconds to achieve direct aberration compensation without error accumulation under strong turbulence.

[0021] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based aberration compensation method for voltage control of a deformable mirror.

[0022] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned deep learning-based aberration compensation method for voltage control of a deformable mirror.

[0023] The beneficial effects of this invention are as follows:

[0024] (1) This invention overcomes the problem of wavefront detection and correction failure of traditional Hartmann wavefront sensors under conditions of light intensity scintillation and phase singularity, and realizes atmospheric turbulence aberration correction under strong turbulence.

[0025] (2) The present invention realizes the direct output of the voltage of the layered deformable mirror from end to end, avoiding the cumulative error brought about by the process of "target wavefront → restored wavefront → deformable mirror surface shape correction" and reducing the system error.

[0026] (3) This invention proposes a combined loss function of pupil light intensity scintillation index and far-field Strell ratio, which realizes physical-guided network optimization and can avoid the uncertainty of multi-deformation image aberration correction coupling and restoration aberration.

[0027] (4) The deep learning network trained by this invention has a millisecond-level computing speed, which meets the real-time requirements of adaptive optics systems and has important application value for high-resolution imaging observation of ground-based telescopes under strong turbulence. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the voltage control method for aberration compensation of a deformable mirror based on deep learning according to the present invention.

[0029] Figure 2 This is a schematic diagram of atmospheric turbulence stratification simulation according to the present invention;

[0030] Figure 3 This is a schematic diagram of the deep learning network based on physical information according to the present invention;

[0031] Figure 4 This is a comparison before and after the present invention uses a trained deep learning network to correct the data in the validation set. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0033] In strong turbulence, pupil surface light intensity flickering drastically reduces wavefront reconstruction accuracy, rendering it unable to detect turbulent aberrations. Previous wavefront reconstruction iterative algorithms required multiple iterations and had limited accuracy. While some wavefront-free methods, including phase difference methods and deep learning, only correct the pupil surface phase, they cannot address phase discontinuities caused by phase singularities or pupil surface light intensity flickering under strong turbulence, resulting in limited performance improvements.

[0034] To address the issues of decreased wavefront reconstruction accuracy and uneven pupil light intensity distribution under strong turbulence, this invention proposes conjugating multiple deformable mirrors at different heights to perform layered correction of atmospheric turbulence, fundamentally resolving the system performance degradation caused by strong turbulence. Furthermore, a physical model of light wave atmospheric transmission and correction is embedded into a deep learning network, constructing a physically based deep learning network that directly outputs control voltages for multiple deformable mirrors, achieving aberration compensation under strong turbulence.

[0035] like Figure 1 As shown in the figure, this embodiment provides a voltage control method for deformable mirrors based on aberration compensation using deep learning, including the following steps:

[0036] Step 1: Use a multi-layer turbulent phase screen to simulate real atmospheric turbulence, and combine it with the angular spectrum diffraction transmission method to establish a layered transmission model of light waves in atmospheric turbulence;

[0037] Step 2: Based on the aforementioned atmospheric turbulence layered transmission model, construct a simulation dataset. Each set of simulation data includes a layered turbulence phase screen, a pupil light intensity image obtained after transmission, and a Shaker-Hartmann image.

[0038] Step 3: Embed the physical model of light wave atmospheric turbulence after layered transmission and correction by multiple deformable mirrors into the deep learning network, build a deep learning network based on physical information, and construct a combined loss function composed of pupil light intensity scintillation index and far-field Strell ratio;

[0039] Step 4: Using the pupil light intensity image and the Shaker-Hartmann image from the simulation dataset as network inputs, train the constructed physical information-based deep learning network. During the training process, substitute the deformable mirror voltages predicted by the network into the physical model to calculate the combined loss function, update the network parameters, and finally output the voltages of multiple deformable mirrors.

[0040] Furthermore, the parameters of the multi-layer turbulent phase screen mentioned in step 1 are obtained based on the atmospheric turbulence stratification theory.

[0041] Furthermore, the deep learning network based on physical information described in step 3 incorporates a physical model of light wave atmospheric transmission and correction, ensuring that network convergence follows physical laws. The model expression is as follows:

[0042] ,

[0043] Where U represents the complex amplitude light field of the pupil plane, U0 represents the incident beam, and the symbol " " represents convolution, and h is the impulse response function, This represents the i-th turbulent phase screen. This represents the transmission distance of the i-th turbulent phase screen. This represents the wavefront loaded by the j-th deformable mirror. Let represent the transmission distance of the wavefront loaded by the j-th deformable mirror, n represent the number of turbulent phase screens, and m represent the number of deformable mirrors.

[0044] Furthermore, the combined loss function described in step 3 ( for:

[0045] ,

[0046] in, Indicates far-field Strelby, The scintillation index, which indicates the intensity of light across the pupil, and These are the corresponding loss weights, set According to the optical diffraction transmission theory, upper-level turbulence mainly affects the light intensity distribution at the pupil plane, thus determining the scintillation index. The far-field Strelby ratio is jointly determined by surface turbulence and upper-level turbulence. Therefore, the combined loss function has two purposes: 1) to ensure that the upper-level deformable mirror corrects for upper-level turbulence and the surface-level deformable mirror corrects for surface turbulence, avoiding the coupling of aberration corrections between multiple deformable mirrors; 2) to avoid the uncertainty of aberration restoration.

[0047] Further, the input in step 4 is the pupil light intensity image and the Shack-Hartmann image, and the output is the control voltage of the layered deformable mirror. Turbulent aberrations at different heights jointly affect the pupil light intensity distribution, and this relationship is non-linear. Furthermore, the Shack-Hartmann image can reflect the turbulent aberrations of the entire atmospheric layer. Therefore, by training a physically based deep learning network, the output voltage of the deformable mirror at different heights can be converged.

[0048] Furthermore, the training process of the deep learning network is as follows: the pupil light intensity image and the Shack-Hartmann image are normalized between 0 and 1 and used as network input. The deep learning network is optimized, and the deformable mirror voltage predicted by the network is substituted into the correction physical model. The loss for each batch of data is calculated using a loss function, and the gradient of the loss with respect to each parameter in the deep learning network is calculated. The parameters are then updated to reduce the loss and eventually converge to a lower level, thus completing the network model training.

[0049] Example:

[0050] In this embodiment, five layers of atmospheric turbulence phase screens were established to simulate real atmospheric turbulence. Each phase screen has a size of 0.8m, and the heights of the five layers are as follows: Layer 1: 0km, Layer 2: 6.8km, Layer 3: 9.8km, Layer 4: 12.1km, Layer 5: 13.9km, etc. Figure 2 As shown. Shaker-Hartmann images and pupil plane intensity images were acquired at the receiver, with the Shaker-Hartmann image array number set to 40×40. The receiver aperture D=0.8m, and the total atmospheric coherence length was set to r0=5cm. A total of 20,000 datasets were generated, each containing 5 layers of atmospheric turbulence phase screens, pupil plane intensity, and Shaker-Hartmann images. The datasets were divided into training, validation, and test sets according to a ratio of 80%, 10%, and 10%, respectively.

[0051] The deep learning network based on physical information constructed in this embodiment is as follows: Figure 3 As shown, the physical model embedded in the network is a 5-layer turbulent phase screen transmission model with 2 deformable mirrors for correction. The parameters of the 5-layer turbulent phase screen are... Figure 2 The parameters for the five-layer atmospheric turbulence simulation are consistent. One deformable mirror is conjugate to 0km to correct surface turbulence aberrations, and another deformable mirror is conjugate to 10km to correct upper-level turbulence aberrations.

[0052] according to Figure 1 The flowchart illustrates the network training process. The network is optimized using the training set. Each time the network outputs a deformable mirror transformer, it is incorporated into the physical model to obtain the far-field image and pupil light intensity after deformable mirror correction. Then, the far-field Strell ratio and the scintillation index of the pupil light intensity are calculated. In this embodiment, the Strell ratio of the far-field image... The ratio of the maximum light intensity of the far-field image after deformable mirror correction to the maximum light intensity of the ideal far-field image is calculated using the following formula:

[0053] ,

[0054] In the above formula, This represents the maximum light intensity of the far-field image after correction by a deformable mirror following atmospheric turbulence propagation of light waves. This represents the maximum light intensity in the ideal far field.

[0055] In this embodiment, the pupillary light intensity scintillation index is calculated using the following formula:

[0056] ,

[0057] In the above formula, I represents the ensemble mean, and I represents the pupillary light intensity. I = |U| 2 .

[0058] The calculated Strell ratio and pupil light intensity flicker index of the far-field image are substituted into the combined loss function for network model training until the loss is reduced and eventually converges to a low level, thus completing the network model training.

[0059] After training, the calibration accuracy of the deep learning network is tested using a validation set. Figure 4 The results are shown for the deep learning network on a validation set. The far-field Strain ratio improved from 0.0783 to 0.6804, and the pupil flicker index decreased from 0.8036 to 0.4349 after correction by the deep learning network. Figure 4 As can be seen, even under conditions of strong turbulence and pupil plane light intensity flicker leading to light loss in the Shack-Hartmann sub-aperture, this invention can still achieve aberration compensation and obtain better far-field results, and the uniformity of pupil plane light intensity is also greatly improved. Therefore, this invention overcomes the limitations of traditional detection and correction methods failing under strong turbulence, and is of great significance for high-resolution imaging observations of ground-based telescopes under strong turbulence.

[0060] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned deep learning-based aberration compensation method for voltage control of a deformable mirror.

[0061] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned deep learning-based aberration compensation method for voltage control of a deformable mirror.

[0062] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A voltage control method for deformable mirrors based on aberration compensation using deep learning, characterized in that, The method is implemented through the following steps: Step 1: Use a multi-layer turbulent phase screen to simulate real atmospheric turbulence, and combine it with the angular spectrum diffraction transmission method to establish a layered transmission model of light waves in atmospheric turbulence; Step 2: Based on the aforementioned atmospheric turbulence layered transmission model, construct a simulation dataset. Each set of simulation data includes a layered turbulence phase screen, a pupil light intensity image obtained after transmission, and a Shaker-Hartmann image. Step 3: Embed the physical model of light wave atmospheric turbulence after layered transmission and correction by multiple deformable mirrors into the deep learning network, build a deep learning network based on physical information, and construct a combined loss function composed of pupil light intensity scintillation index and far-field Strell ratio; Step 4: Using the pupil light intensity image and the Shaker-Hartmann image from the simulation dataset as network inputs, train the constructed physical information-based deep learning network. During the training process, substitute the deformable mirror voltages predicted by the network into the physical model, calculate the combined loss function to update the network parameters, and finally output the voltages of multiple deformable mirrors.

2. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, The parameters of the multi-layer turbulent phase screen mentioned in step 1 are obtained based on the atmospheric turbulence stratification theory.

3. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, The physical model embedded in the deep learning network in step 3 is used to describe the process of wavefront distortion correction after the beam passes through the turbulent phase screen for layered diffraction and transmission, and then is loaded by multiple deformable mirrors. During the process, the pupil light intensity and far-field image are obtained. The number of turbulent phase screens and the number of deformable mirrors are set according to the layered correction requirements.

4. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, The combined loss function described in step 3 is composed of a weighted sum of a pupil intensity scintillation index term and a far-field Strell ratio term; wherein the far-field Strell ratio is used to evaluate the quality of the far-field spot after correction by the deformable mirror, and the pupil intensity scintillation index is used to evaluate the uniformity of pupil intensity after correction by the deformable mirror, and the sum of the weight coefficients of the two terms is one.

5. The method for voltage control of a deformable mirror based on aberration compensation according to claim 4, characterized in that, The combined loss function is used to correct high-level turbulence with high-level deformable mirrors and correct surface turbulence with surface deformable mirrors, avoiding the uncertainty of multi-deformable mirror aberration correction coupling and restoration aberration.

6. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, Turbulent aberrations at different altitudes collectively affect the pupil light intensity distribution in a nonlinear manner, and the Shaker-Hartmann image is used to reflect the turbulent aberrations of the entire atmospheric layer.

7. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, The training process described in step 4 is as follows: after normalizing the pupil light intensity image and the Shaker-Hartmann image, the network input is used; the deformable mirror voltage predicted by the network is substituted into the physical model; the loss of each batch of data is calculated through the combined loss function; the gradient of the loss with respect to each parameter in the network is calculated and the parameters are updated until the loss converges.

8. The method for voltage control of a deformable mirror based on aberration compensation according to claim 1, characterized in that, The trained, physically-based deep learning network is used to output the voltages of multiple deformable mirrors in milliseconds to achieve direct aberration compensation without error accumulation under strong turbulence.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based aberration compensation voltage control method for deformable mirrors as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the deep learning-based aberration compensation voltage control method for deformable mirrors as described in any one of claims 1-8.

Images