High-fidelity dynamic turbulence phase screen rapid generation method based on digital twinborn technology

Abstract

The invention discloses a high-fidelity dynamic turbulence phase screen rapid generation method based on a digital twinning technology, belongs to the technical field of atmospheric turbulence simulation, and solves the problem that the physical property interpretability of a model is insufficient due to the fact that the training process of an existing neural network dynamic turbulence generation model directly depends on initial phase screen characteristics. The method comprises the following steps: decomposing an initial phase screen based on a turbulence characteristic decomposition and extraction module Rand-Net; capturing spatial structure characteristics of the phase screen by adopting a convolutional network CNN, carrying out explicit correlation on a wind speed vector by utilizing a space-time convolution kernel, modeling a phase screen time evolution process driven by the wind speed, and outputting a phase screen at the next moment; based on the digital twinning technology, data such as a physical model and operation history are fully utilized, a phase screen generation frame with both physical rule controllability and calculation efficiency is constructed, and meanwhile, a reliable technical means is provided for high-precision modeling of a self-adaptive optical system through rapid generation of the high-fidelity dynamic turbulence phase screen.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric turbulence simulation technology, specifically relating to a method for rapid generation of high-fidelity dynamic turbulence phase screens based on digital twin technology. Background Technology

[0002] Adaptive Optics (AO) technology, as a core means to solve the effects of atmospheric turbulence, has played a crucial role in high-resolution observations with large-aperture telescopes and high-speed satellite-to-ground laser communication. However, high-fidelity dynamic characterization of turbulence models remains one of the key technologies for improving the simulation accuracy of AO systems. Traditional turbulence phase screen modeling methods are mainly divided into two categories: one is a static model based on Zernike polynomials and Taylor's frozen turbulence assumption, which expands the phase screen through spatiotemporal translation, but its finite size and boundary effects lead to long-term simulation distortion; the other is the extrapolation method, which establishes a dynamic correlation model by mining the spatiotemporal evolution law of the phase field. Although this improves the efficiency of sequence generation, it is still limited by the prior assumptions of the initial statistical characteristics.

[0003] In recent years, researchers have attempted to improve modeling results by integrating statistical properties of turbulence with numerical methods. In 2024, Lu Benchu's team proposed a hybrid modeling method, reconstructing the Zernike coefficient covariance matrix using an improved von Kármán spectrum and generating a dynamic phase screen using inverse Fourier transform. While this method achieved progress in adaptability to internal and external scale effects and complex wind fields, it still relied on the presuppositions of the theoretical power spectrum model. In 2025, Cao Zhongyu's team achieved high-precision simulations based on atmospheric coherence length and the von Kármán model, utilizing pixel displacement correlation. Simultaneously, Fu Haiyang's team improved spatiotemporal resolution using Kolmogorov power spectrum and scale transformation interpolation techniques, but these approaches are essentially numerical optimizations of classical turbulence models, struggling to overcome the limitations of theoretical assumptions.

[0004] The methods described above are all based on forward modeling and calculation of turbulent statistical properties, requiring prior assumptions about power spectrum models, etc. However, in real dynamic turbulence scenarios, the spatiotemporal random fluctuations of atmospheric refractive index lead to wavefront distortion exhibiting strong nonlinear and non-stationary characteristics. When actual atmospheric turbulence parameters deviate from the assumptions of standard Kolmogorov power spectrum models, manual intervention is required to extract statistical features from measured data, making it difficult for dynamic turbulence phase screen generation methods based on statistical models to balance the accuracy of turbulence physical characterization with dynamic generation efficiency. To overcome this bottleneck, researchers utilized deep neural networks to effectively capture the spatiotemporal correlation characteristics of phase screens directly from measured data, thereby rapidly generating infinitely long dynamic phase screen sequences that conform to the statistical laws of actual turbulence, demonstrating the feasibility of combining turbulence spatiotemporal feature extraction with neural networks. However, the end-to-end dynamic turbulence rapid generation model based on neural networks relies directly on real turbulence features during its training process. It has not yet established an explicit mapping relationship between the network's hidden layer feature space and statistical parameters such as atmospheric coherence length (r0), turbulence outer scale (L0), and inner scale (l0). This makes it difficult to actively constrain and control the turbulence characteristics of the generated phase screen through statistical parameters. Therefore, it suffers from insufficient physical interpretability and coupling of physical parameters when constructing diverse high-fidelity dynamic turbulence datasets. Therefore, we propose a high-fidelity dynamic turbulence phase screen rapid generation method based on digital twin technology. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a high-fidelity dynamic turbulence phase screen rapid generation method based on digital twin technology. This method solves the problem that although existing neural network dynamic turbulence generation models can quickly generate phase screen sequences that conform to statistical laws based on measured data, their training process directly depends on the initial phase screen features, resulting in insufficient interpretability of the model's physical properties.

[0006] This invention is implemented as follows: a method for rapid generation of high-fidelity dynamic turbulent phase screens based on digital twin technology, the method comprising:

[0007] S10, Construct the physical parameter coupled turbulence feature decomposition and extraction module Rand-Net, and decompose the initial phase screen based on the turbulence feature decomposition and extraction module Rand-Net;

[0008] S20 uses a convolutional network (CNN) to capture the spatial structure features of the phase screen;

[0009] S30 is used to obtain the spatial structure features of the phase screen, construct the conditional phase screen generation module Phs-Net based on the conditional generative adversarial network, and generate a phase screen set through the phase screen generation module Phs-Net.

[0010] S40, design the dynamic turbulence evolution module Next-Net, which uses spatiotemporal convolution kernels to explicitly correlate wind speed vectors, model the wind speed-driven temporal evolution process of the phase screen, and output the phase screen at the next moment;

[0011] S50 designs physical constraints and training strategies, including parameter closed-loop verification constraints, to jointly optimize generation quality, physical law compliance, and turbulence parameter controllability.

[0012] Preferably, when decomposing the initial phase screen based on the Rand-Net turbulence feature decomposition and extraction module, the initial real phase screen is first decomposed... Decomposed into deterministic physical characteristics and turbulent random fluctuation vector .

[0013] Preferably, when using a convolutional neural network (CNN) to capture the spatial structure features of the phase screen, the spatial structure features of the phase screen are represented as follows: Physical characteristics are transmitted through a fully connected network MLP. Mapped to a statistical parameter condition vector;

[0014] The statistical parameter condition vector is represented as follows: , Represents statistical parameters Statistical parameters are simultaneously input into the Rand-Net input to the turbulence feature decomposition and extraction module. Output and Furthermore, regarding physical characteristics and turbulent random fluctuation vector Perform explicit decoupling of features.

[0015] Preferably, regarding physical characteristics and turbulent random fluctuation vector When performing explicit feature decoupling, the explicit feature decoupling formula is expressed as:

[0016]

[0017] By constructing a parametric regression subnetwork to constrain the mapping relationship between physical characteristics and statistical parameters, the parametric regression subnetwork... The loss function is expressed as:

[0018] .

[0019] Preferably, the conditional phase screen generation module Phs-Net uses a conditional generative adversarial network to process the spatial structure features of the phase screen. Phs-Net includes a condition generator and a discriminator. Phs-Net uses statistical parameters... and random vectors Concatenation is used as a conditional input to generate a phase screen set. .

[0020] Preferably, the discriminator is designed with a multi-scale structure. Perform phase-based pixel-level similarity determination. Perform power spectral density discrimination. Statistical parametric regression is performed, and the generator responds to the discriminator using the following loss function:

[0021] .

[0022] Preferably, the dynamic turbulence evolution module Next-Net is used to acquire the phase screen set. Phase screen set Current phase screen Combined with wind speed estimation and random fluctuation vector As input, output the phase screen at the next time step. ;

[0023] Among them, when explicitly associating wind speed vectors using spatiotemporal convolution kernels, the spatiotemporal convolution kernels... The weights are explicitly related to the wind speed vector, and the spatiotemporal convolution kernel... Defined as:

[0024] .

[0025] Preferably, the parameter closed-loop validation constraint is based on the parameter regression subnetwork definition, and the parameter closed-loop validation constraint is defined as follows:

[0026] .

[0027] Compared with the prior art, the embodiments of this application have the following main advantages:

[0028] This invention, based on digital twin technology, fully utilizes physical models and operational history data to construct a phase screen generation framework that combines controllability of physical laws with computational efficiency. Furthermore, by rapidly generating high-fidelity dynamic turbulent phase screens, this invention provides a reliable technical means for high-precision modeling of adaptive optics (AO) systems. Because this invention can accurately simulate the physical characteristics of turbulence and can be precisely controlled according to actual needs, it can provide a near-real-world testing platform for control algorithms and wavefront reconstruction algorithms involved in AO systems. In addition, based on the accurate wavefront distortion information obtained from the simulation, the AO system can optimize the control strategy of the wavefront corrector, effectively improving the imaging quality of the optical system. Attached Figure Description

[0029] Figure 1 A schematic diagram illustrating the implementation process of a method for rapidly generating high-fidelity dynamic turbulent phase screens based on digital twin technology is shown. Detailed Implementation

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] To address the issue that while existing neural network dynamic turbulence generation models can rapidly generate statistically consistent phase screen sequences based on measured data, their training process directly relies on initial phase screen features, leading to insufficient interpretability of the model's physical properties, we propose a high-fidelity dynamic turbulence phase screen rapid generation method based on digital twin technology. In short, the method first constructs a turbulence feature decomposition and extraction module, Rand-Net, coupled with physical parameters. Based on Rand-Net, the initial phase screen is decomposed. Then, a convolutional network (CNN) is used to capture the spatial structure features of the phase screen. Finally, a conditional generation method is constructed. The Phs-Net module, a conditional phase screen generation module for adversarial networks, generates a set of phase screens. Simultaneously, a Next-Net module for dynamic turbulence evolution is designed, which explicitly correlates wind speed vectors using spatiotemporal convolution kernels to model the wind-driven temporal evolution of the phase screens and outputs the phase screen at the next moment. Based on digital twin technology, this invention fully utilizes physical models and historical data to construct a phase screen generation framework that combines controllability of physical laws with computational efficiency. Furthermore, this invention provides a reliable technical means for high-precision modeling of AO systems through the rapid generation of high-fidelity dynamic turbulence phase screens. Because this invention can accurately simulate the physical characteristics of turbulence and can be precisely controlled according to actual needs, it can provide a near-real-world testing platform for control algorithms and wavefront reconstruction algorithms involved in AO systems. In addition, based on the accurate wavefront distortion information obtained from the simulation, the AO system can optimize the control strategy of the wavefront corrector, effectively improving the imaging quality of the optical system.

[0033] This invention provides a method for rapid generation of high-fidelity dynamic turbulent phase screens based on digital twin technology. Figure 1 A schematic diagram illustrating the implementation process of a high-fidelity dynamic turbulence phase screen rapid generation method based on digital twin technology is shown. The method specifically includes:

[0034] S10, Construct the physical parameter coupled turbulence feature decomposition and extraction module Rand-Net, and decompose the initial phase screen based on the turbulence feature decomposition and extraction module Rand-Net;

[0035] It should be noted that when decomposing the initial phase screen based on the Rand-Net turbulence feature decomposition and extraction module, the initial real phase screen is first decomposed. Decomposed into deterministic physical characteristics and turbulent random fluctuation vector .

[0036] Meanwhile, the turbulence feature decomposition and extraction module Rand-Net can also be used to decompose the simulated phase screen, the physical features of which include, but are not limited to, atmospheric coherence length r0, atmospheric external scale L0, and atmospheric internal scale l0 parameters.

[0037] S20 uses a convolutional network (CNN) to capture the spatial structure features of the phase screen;

[0038] Wherein, when using a convolutional network (CNN) to capture the spatial structure features of the phase screen, the spatial structure features of the phase screen are represented as follows: Physical characteristics are transmitted through a fully connected network MLP. Mapped to a statistical parameter condition vector;

[0039] The statistical parameter condition vector is represented as follows: , Represents statistical parameters Statistical parameters are simultaneously input into the Rand-Net input to the turbulence feature decomposition and extraction module. Output and Furthermore, regarding physical characteristics and turbulent random fluctuation vector Explicit decoupling of features is performed, thus laying the foundation for parameterized generation.

[0040] For physical characteristics and turbulent random fluctuation vector When performing explicit feature decoupling, the explicit feature decoupling formula is expressed as:

[0041]

[0042] To ensure at the same time It can accurately encode statistical parameters and network feature space without deviating from physical interpretability. By constructing a parametric regression subnetwork, it constrains the mapping relationship between physical features and statistical parameters. The loss function is expressed as:

[0043] .

[0044] S30 is used to obtain the spatial structure features of the phase screen, construct the conditional phase screen generation module Phs-Net based on the conditional generative adversarial network, and generate a phase screen set through the phase screen generation module Phs-Net.

[0045] The conditional phase screen generation module Phs-Net employs a conditional generative adversarial network to process the spatial structural features of the phase screen. Phs-Net includes a condition generator and a discriminator. Phs-Net can be understood as the inverse process of Rand-Net. The conditional phase screen generation module Phs-Net uses statistical parameters... and random vectors Concatenation is used as a conditional input to generate a phase screen set. .

[0046] In this embodiment, the generator introduces a conditional vector at each layer through conditional batch normalization (CBN) processing. Thus ensuring It will be based on the input The phase screen generation process, which involves adjustments to achieve control over turbulent physical parameters, is represented as follows:

[0047]

[0048] in, (mean) and (Standard deviation) is a normalization parameter. and for The function.

[0049] In this embodiment of the invention, the multi-scale generation constraint first constrains the phase pixel-level similarity through basic reconstruction:

[0050]

[0051] This will then generate the power spectral density of the phase screen. Limited to the theoretical framework of turbulence statistics (e.g., Kolmogorov power spectrum or power spectrum fitted by actual measurements). To ensure the correct frequency domain characteristics of the generated turbulence, the loss function used to constrain the power spectral density of the generated phase screen is defined as follows:

[0052]

[0053] in, and Representing coordinates in frequency space, used to determine the power spectral density values ​​at different frequency locations. M and N represent... The size of the frequency space refers to the range of the frequency spectrum. It is a weighted matrix used to adjust the weights of different frequency position differences based on the power spectral density differences.

[0054] The discriminator is designed with a multi-scale structure. Perform phase-based pixel-level similarity determination. Perform power spectral density discrimination. Statistical parametric regression is performed, and the generator responds to the discriminator using the following loss function:

[0055] .

[0056] S40, design the dynamic turbulence evolution module Next-Net, which uses spatiotemporal convolution kernels to explicitly correlate wind speed vectors, models the wind speed-driven temporal evolution of the phase screen, and outputs the phase screen at the next moment. It should be noted that as the wind speed increases ( (Increase) leading to convolution kernel If the size is increased, the phase screen will displace faster in space, thereby achieving the advection effect that simulates wind-driven turbulence;

[0057] In this embodiment of the invention, the Next-Net dynamic turbulence evolution module is used to acquire the phase screen set. Phase screen set Current phase screen Combined with wind speed estimation and random fluctuation vector As input, output the phase screen at the next time step. ;

[0058] Among them, when explicitly associating wind speed vectors using spatiotemporal convolution kernels, the spatiotemporal convolution kernels... The weights are explicitly related to the wind speed vector, and the spatiotemporal convolution kernel... Defined as:

[0059]

[0060] S50 designs physical constraints and training strategies, including parameter closed-loop verification constraints, to jointly optimize generation quality, physical law compliance, and turbulence parameter controllability.

[0061] It should be noted that, in order to jointly optimize the generation quality, compliance with physical laws, and controllability of turbulence parameters, in addition to ensuring the correct frequency domain characteristics of the generated turbulence, the coding features of the phase screen are forcibly generated. It can reverse regress to obtain the original statistical parameters. This ensures the reliability of parameter control. Combined with the parameter regression subnetwork definition, parameter closed-loop verification constraints are defined as follows:

[0062]

[0063] Meanwhile, the phase screen generation framework training process first establishes basic physical relationships (pre-training), then improves generation quality through adversarial learning, while a multi-scale discriminator ensures that physical laws at different levels are satisfied (adversarial fine-tuning). The pre-training stage of the phase screen generation framework first uses a frozen regression network. This means training only the encoder (Rand-Net) and generator (Phs-Net) to avoid regression errors interfering with feature decoupling. This then unfreezes the code. ,pass Joint optimization.

[0064] In this embodiment, a phase screen generation framework that combines controllability of physical laws with computational efficiency is constructed. This framework consists of a turbulence feature decomposition and extraction module (Rand-Net), a convolutional network (CNN), a phase screen generation module (Phs-Net), and a dynamic turbulence evolution module (Next-Net). The collaborative operation of these modules avoids the reliance on repeatedly calculating historical data, significantly improving the efficiency of dynamic phase screen generation. Furthermore, its controllable physical parameters provide the AO system with diverse and high-precision turbulence datasets, facilitating the optimization and validation of the wavefront correction algorithm.

[0065] This invention, based on digital twin technology, fully utilizes physical models and operational history data to construct a phase screen generation framework that combines controllability of physical laws with computational efficiency. Furthermore, by rapidly generating high-fidelity dynamic turbulent phase screens, this invention provides a reliable technical means for high-precision modeling of AO systems. Because this invention can accurately simulate the physical characteristics of turbulence and can be precisely controlled according to actual needs, it can provide a near-real-world testing platform for control algorithms and wavefront reconstruction algorithms involved in AO systems. In addition, based on the accurate wavefront distortion information obtained from the simulation, the AO system can optimize the control strategy of the wavefront corrector, effectively improving the imaging quality of the optical system.

[0066] In summary, this invention provides a rapid method for generating high-fidelity dynamic turbulent phase screens based on digital twin technology. This invention, based on digital twin technology, fully utilizes physical models and operational history data to construct a phase screen generation framework that combines controllability of physical laws with computational efficiency. Furthermore, the rapid generation of high-fidelity dynamic turbulent phase screens provides a reliable technical means for high-precision modeling of optical oscillators (AO) systems. Because this invention can accurately simulate the physical characteristics of turbulence and can be precisely controlled according to actual needs, it can provide a near-real-world testing platform for control algorithms and wavefront reconstruction algorithms involved in AO systems. In addition, based on the accurate wavefront distortion information obtained from the simulation, the AO system can optimize the control strategy of the wavefront corrector, effectively improving the imaging quality of the optical system.

[0067] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for rapid generation of high-fidelity dynamic turbulent phase screens based on digital twin technology, characterized in that, The method includes: S10, Construct the physical parameter coupled turbulence feature decomposition and extraction module Rand-Net, and decompose the initial phase screen based on the turbulence feature decomposition and extraction module Rand-Net; S20 uses a convolutional network (CNN) to capture the spatial structure features of the phase screen; S30 is used to obtain the spatial structure features of the phase screen, construct the conditional phase screen generation module Phs-Net based on the conditional generative adversarial network, and generate a phase screen set through the phase screen generation module Phs-Net. S40, design the dynamic turbulence evolution module Next-Net, which uses spatiotemporal convolution kernels to explicitly correlate wind speed vectors, model the wind speed-driven temporal evolution process of the phase screen, and output the phase screen at the next moment; S50 designs physical constraints and training strategies, including parameter closed-loop verification constraints, to jointly optimize generation quality, physical law compliance, and turbulence parameter controllability.

2. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 1, characterized in that: When decomposing the initial phase screen based on the Rand-Net turbulence feature decomposition and extraction module, the initial real phase screen is first decomposed... Decomposed into deterministic physical characteristics and turbulent random fluctuation vector .

3. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 2, characterized in that: When using a convolutional neural network (CNN) to capture the spatial structure features of the phase screen, the spatial structure features of the phase screen are represented as follows: Physical characteristics are transmitted through a fully connected network MLP. Mapped to a statistical parameter condition vector; The statistical parameter condition vector is represented as follows: , Represents statistical parameters In the Rand-Net input to the turbulence feature decomposition and extraction module, statistical parameters are input simultaneously. Output and Furthermore, regarding physical characteristics and turbulent random fluctuation vector Perform explicit decoupling of features.

4. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 3, characterized in that: For physical characteristics and turbulent random fluctuation vector When performing explicit feature decoupling, the explicit feature decoupling formula is expressed as: By constructing a parametric regression subnetwork to constrain the mapping relationship between physical characteristics and statistical parameters, the parametric regression subnetwork... The loss function is expressed as: 。 5. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 2, characterized in that: The conditional phase screen generation module Phs-Net uses a conditional generative adversarial network to process the spatial structure features of the phase screen. Phs-Net includes a condition generator and a discriminator. Phs-Net uses statistical parameters... and random vectors Using concatenation as a conditional input, a phase screen set is generated. .

6. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 5, characterized in that: The discriminator is designed with a multi-scale structure. Perform phase-based pixel-level similarity determination. Perform power spectral density discrimination. Statistical parametric regression is performed, and the generator responds to the discriminator using the following loss function: 。 7. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 5, characterized in that: The Next-Net dynamic turbulence evolution module is used to acquire the phase screen set. Phase screen set Current phase screen Combined with wind speed estimation and random fluctuation vector As input, output the phase screen at the next time step. ; Among them, when explicitly associating wind speed vectors using spatiotemporal convolution kernels, the spatiotemporal convolution kernels... The weights are explicitly related to the wind speed vector, and the spatiotemporal convolution kernel... Defined as: 。 8. The method for rapid generation of high-fidelity dynamic turbulent phase screen based on digital twin technology as described in claim 7, characterized in that: The parameter closed-loop validation constraint is based on the parameter regression subnetwork definition, and the parameter closed-loop validation constraint is defined as follows: 。

Images