Satellite thermal stress rapid prediction method and system based on domain decomposition

Abstract

The invention provides a satellite thermal stress rapid prediction method and system based on domain decomposition, and the method comprises the steps: constructing an unstructured grid thermal stress data set composed of temperature and stress data pairs, and employing local grid refinement to obtain a high-stress gradient region with a higher resolution; constructing a K-D tree model optimized by Monte Carlo tree search, and adaptively decomposing the domain into a high-gradient sub-domain and a low-gradient sub-domain according to the coordinates of the point cloud and the thermal stress gradient; constructing a high-precision physical field proxy model of a plurality of neural operators, and respectively applying GeoFNO and NU-FNO in a high stress gradient sub-domain and a low stress gradient sub-domain; in a high stress gradient sub-domain, GeoFNO is adopted to directly learn on non-structural data, in a low stress gradient sub-domain, NU-FNO interpolates a point cloud into a uniform grid, and then based on fast Fourier transform learning mapping, thermal stress under various thermal loads and complex geometrical shapes is predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of satellite thermal stress prediction, and in particular relates to a method and system for rapid prediction of satellite thermal stress based on domain decomposition. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Rapid thermal stress prediction is critical to satellite thermal design. Satellites operate in harsh environments in orbit, periodically subjected to alternating high and low temperatures from outer space, resulting in significant temperature gradients. Furthermore, due to the high power density of satellite components, they inevitably generate significant amounts of heat during operation. The combination of severe external temperature differences and internal heat sources can easily lead to thermal deformation, thermal buckling, or thermally induced vibration, posing significant challenges to structural stability, equipment reliability, and mission accuracy. Traditional thermal stress analysis methods primarily rely on solving partial differential equations to implement numerical methods such as finite differences and finite elements. These traditional algorithms typically offer high accuracy but struggle to achieve rapid response.

[0004] Establishing surrogate models is a common approach to achieve a compromise between computational accuracy and cost, improving the efficiency of prediction tasks. Traditional surrogate modeling methods, including polynomial-based response surfaces, support vector machine regression, radial basis functions, and Kriging interpolation, all face the challenge of the "curse of dimensionality," where the high dimensionality of inputs and outputs makes modeling exponentially more difficult. Deep learning models are considered advantageous in processing high-dimensional data. Due to the universal approximation properties and computational efficiency of neural networks, data-driven deep learning surrogate modeling techniques offer new avenues for constructing high-precision physics surrogate models. For example, current research uses models such as graph neural networks and convolutional neural networks to learn the mapping relationship between temperature and stress fields. However, in practical applications with complex geometries and unstructured data, uniform grid-based methods are no longer applicable. Meshless techniques can flexibly adapt to complex geometries. Therefore, learning the complex stress characteristics of each region, overcoming the limitations of complex geometries and unstructured data, and enabling prediction of drastically changing physical fields remain key challenges in the construction of high-precision physics surrogate models.

[0005] Satellite thermal stress prediction tasks usually face constraints such as complex models, limited computing resources, and strong real-time performance. Traditional numerical methods such as finite element methods are difficult to meet the requirements. Data-driven deep learning proxy models have gradually become the mainstream technology for building high-precision physical field proxy models.

[0006] Existing rapid thermal stress prediction methods for neural network models construct a mapping relationship between temperature fields and stress fields. These neural network models are trained through supervised learning, thereby establishing an end-to-end, high-precision proxy model for the satellite's temperature-stress transition, enabling rapid thermal stress prediction. However, in applications with complex geometries, particularly satellite panels with screw holes, stress concentrations occur near the holes. A single neural operator struggles to capture the complex stress characteristics of these different regions, resulting in low prediction accuracy for the resulting proxy model.

[0007] Therefore, in practical applications involving complex satellite panels and unstructured data, uniform mesh-based methods are no longer applicable. Because stress concentration in the holes of satellite panels with screw holes leads to significant differences in stress distribution across different regions, a single neural operator based on meshless technology struggles to learn the complex temperature-stress mapping relationship, resulting in low prediction accuracy for the resulting proxy model. Summary of the Invention

[0008] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a method for rapid prediction of satellite thermal stress based on domain decomposition, which can achieve rapid prediction of thermal stress under complex stress distribution. It has high efficiency, high accuracy and strong versatility, thereby providing guidance and reference for better solving the construction of high-precision proxy models in complex satellite configuration scenarios.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] In a first aspect, a method for rapid prediction of satellite thermal stress based on domain decomposition is disclosed, comprising:

[0011] Construct an unstructured grid thermal stress dataset consisting of temperature and stress data pairs, and use local grid refinement to obtain higher resolution in high stress gradient areas;

[0012] A KD tree model optimized by Monte Carlo tree search is constructed to adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient;

[0013] Construct a high-precision physical field proxy model of multiple neural operators, and apply Geo-FNO and NU-FNO in the high stress gradient subdomain and low stress gradient subdomain respectively;

[0014] In the high stress gradient subdomain, Geo-FNO is used to learn directly on unstructured data. In the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid and then learns a mapping based on fast Fourier transform to predict thermal stresses under various thermal loads and complex geometries.

[0015] As a further technical solution, the process of constructing an unstructured grid thermal stress dataset consisting of temperature and stress data pairs is as follows:

[0016] Input temperature field, which is a sample in Gaussian random field;

[0017] When adjusting the mean and covariance of the Gaussian random field, the roughness of the temperature field changes;

[0018] No-displacement boundary conditions are applied to the edges of the inner hole, and free boundary conditions are applied to the rectangular boundaries of the main plate;

[0019] The thermal stress field is calculated using the finite element method.

[0020] As a further technical solution, a KD tree model optimized by Monte Carlo tree search is constructed to adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient. Specifically, the following are the steps:

[0021] The optimization goal of domain decomposition is to minimize the sum of squares of stress gradients in each subdomain;

[0022] Each node of the Monte Carlo tree represents a KD tree and can take an action to move to its child node;

[0023] The action of the Monte Carlo Tree Search agent consists of three parts, including selecting the subdomain D * , choose a dimension from d * and using a hyperplane x d =b * Divide the selected subdomain D * .

[0024] As a further technical solution, for regional decomposition, the subdomain division results at the end of each agent simulation are evaluated. (0) TSS in and TSS in all subdomains i The sum defines the final reward.

[0025] As a further technical solution, the optimization objective of regional decomposition is described as:

[0026]

[0027] Where, The TSS consists of non-intersecting rectangular subdomains. i Indicates subdomain D (i) The sum of the squares of the differences between the gradient of each point in and the average gradient is as follows:

[0028]

[0029] Where M (i)Indicates D (i) The number of midpoints, Indicates subdomain D (i) The magnitude of the thermal stress gradient. Indicates subdomain D (i) average value of the thermal stress gradient;

[0030] Equation (2) approximates the solution to this problem via the Monte Carlo tree search algorithm.

[0031] As a further technical solution, a high-precision physical field proxy model of multiple neural operators is constructed, and Geo-FNO and NU-FNO are applied in the high stress gradient subdomain and low stress gradient subdomain respectively. The detailed processing process is as follows:

[0032] Data preprocessing: First, in the high stress gradient subdomain In the example, the physical grid T is connected through a fully connected network. h,P Deformed to an adaptive grid T on the computational domain h,C ; In the low stress gradient region , input point cloud Interpolate to the input subdomain grid T l,C ;

[0033] Data Dimensionality Upscaling: Adaptive Grid T h,C Temperature embedding in Transformed into hidden layer embedding through a fully connected network, in the subdomain grid T l,C Medium temperature embedding It is also converted into hidden layer embedding through a fully connected network;

[0034] Information transmission: Grid T h,C and T l,C The hidden layer embedding in F is respectively transformed by geometric Fourier transform. a 、 and Fourier transform F, F -1 Pass L wheel;

[0035] Data dimensionality reduction: hidden layer embeddings are transformed into computational grid T through a fully connected network h,C and T l,C middle;

[0036] Data post-processing: Calculation grid T h,C and T l,C Deform and interpolate back to the physical mesh T separately h,P and T l,P The final thermal stress field is obtained.

[0037] In a second aspect, a satellite thermal stress rapid prediction system based on domain decomposition is disclosed, comprising:

[0038] The unstructured grid thermal stress dataset construction module is configured to: construct an unstructured grid thermal stress dataset consisting of temperature and stress data pairs, and use local grid refinement to obtain higher resolution high stress gradient areas;

[0039] A KD tree model building module is configured to: build a KD tree model optimized by Monte Carlo tree search, and adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient;

[0040] A high-precision physics proxy model construction module is configured to: construct a high-precision physics proxy model of multiple neural operators, apply Geo-FNO and NU-FNO in the high stress gradient subdomain and the low stress gradient subdomain respectively;

[0041] The thermal stress prediction module is configured as follows: in the high stress gradient subdomain, Geo-FNO is used to learn directly on unstructured data; in the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid and then learns the mapping based on fast Fourier transform to predict thermal stress under various thermal loads and complex geometries.

[0042] One or more of the above technical solutions have the following beneficial effects:

[0043] The technical solution of the present invention proposes a satellite thermal stress rapid prediction technology based on domain decomposition. The calculation domain is divided into multiple subdomains according to the stress distribution characteristics, and the stress characteristics of each subdomain are learned by multiple neural operators. This overcomes the limitations of complex geometric shapes and unstructured data application scenarios, realizes rapid prediction of thermal stress under complex stress fields, and forms a set of high-efficiency, high-precision, and highly versatile thermal stress rapid prediction methods, thereby providing guidance and reference for better solving the problem of constructing high-precision proxy models in complex stress scenarios.

[0044] The distribution characteristic of the technical solution of the present invention is that high stress gradients are concentrated in the area surrounding the hole, while the remaining areas exhibit low stress gradients. The computational domain is the satellite cabin panel. The advantage of regional decomposition based on stress distribution characteristics is that different prediction models are used for sub-regions with different stress distribution characteristics, balancing prediction accuracy and calculation time.

[0045] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0047] Figure 1 This is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0049] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0050] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0051] Existing research on satellite thermal stresses focuses on simple rectangular plates without screw holes. It uses models based on uniform grids, or structured data, to predict thermal stress. This is not suitable for the satellite plates with screw holes studied in the present invention. The present invention uses a model based on point clouds, or unstructured data, to predict thermal stress, which is more relevant to real-world applications. In real-world applications, temperature measurements using sensors typically generate unstructured data.

[0052] The overall idea proposed by the present invention is:

[0053] For applications involving satellite panels with complex geometries and unstructured data, the GEO-FNO neural operator can directly process unstructured data, avoiding high interpolation errors, while the NU-FNO operator can efficiently process low-stress gradient regions, achieving high-precision thermal stress prediction. Combining the characteristics of these two neural operators, they are applied to the construction of high-precision physical field proxy models, enabling rapid thermal stress prediction of satellite panels with complex geometries. This results in a highly efficient, accurate, and versatile method for rapid thermal stress prediction. The key to this method lies in its initial application of the domain decomposition method to high-precision physical field proxy models, overcoming their limitations in application scenarios. This results in the design and implementation of a highly efficient and versatile method for rapid thermal stress prediction.

[0054] Example 1

[0055] This embodiment discloses a method for rapid prediction of satellite thermal stress based on domain decomposition. First, the mapping relationship is introduced:

[0056] The rapid prediction of thermal stress by building a proxy model for neural networks is usually done in an end-to-end manner. The mapping from temperature field to physical field is trained as follows:

[0057]

[0058] Where: is the mapping from the temperature field T to the stress field σ. This method uses mean absolute error and standard deviation of error as performance metrics. During the training phase, the model is trained using mean square error loss. After learning an operator that maps the temperature field to the stress field, the model can be applied to satellite component thermal design and layout during the satellite design phase, as well as to monitoring the operating status of satellite components during the on-orbit phase.

[0059] Aiming at the application scenarios of satellite panels with complex geometries and unstructured data, a fast prediction method of satellite thermal stress based on domain decomposition is proposed.

[0060] The model in this invention can be directly applied to unstructured data, whereas previous research has only been able to apply to structured data. This is because the GEO-FNO model uses learning in the computational domain to transform unstructured data into structured data, while the NU-FNO model uses interpolation to transform unstructured data into structured data before learning.

[0061] Unstructured meshes are more suitable for satellite panels with screw holes than structured data, which is close to engineering application scenarios.

[0062] In the technical solution of this implementation patent, first, an unstructured grid thermal stress data set consisting of temperature and stress data pairs is proposed, and local grid refinement is used to make the resolution of high stress gradient areas higher, and the unstructured grid can adapt to complex geometric shapes. Secondly, a KD tree model optimized by Monte Carlo tree search is constructed to adaptively decompose the domain into high gradient subdomains and low gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient. Finally, in the high stress gradient subdomain, Geo-FNO is used to learn directly on the unstructured data, avoiding large interpolation errors. In the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid, and then efficiently and accurately learns the mapping based on the fast Fourier transform method. This highlights their complementarity. Therefore, thermal stresses under various thermal loads and complex geometries can be predicted quickly and accurately. Figure 1 The process of the rapid thermal stress prediction method based on domain decomposition is presented and the method is described in detail below. It specifically includes the following steps:

[0063] Step 1: Construct an unstructured grid thermal stress dataset.

[0064] Unstructured meshes are used to flexibly handle complex geometries. The presence of holes causes intense thermal stresses near the holes, resulting in stress concentrations. To accurately calculate the thermal stresses in these areas, local mesh refinement is used to achieve finer resolution.

[0065] The input temperature field is a sample of a Gaussian random field. Adjusting the mean and covariance of the Gaussian random field changes the roughness of the temperature field. A no-displacement boundary condition is applied to the inner hole edge, and a free boundary condition is applied to the rectangular boundary of the mainboard. The thermal stress field is calculated using the finite element method.

[0066] In this implementation example, the steps of constructing a dataset include:

[0067] (1) Draw an unstructured grid and then refine the grid around the hole to obtain more accurate simulation results in the finite element method.

[0068] (2) By setting the three parameter values ​​of the Gaussian random temperature field: mean, variance, and smoothness, an m×n temperature field can be obtained. m and n are the number of grids in the two directions of the temperature field set by oneself. In this embodiment, they are set to 400 and 200. Then, the temperature values ​​at the unstructured grid points are obtained by nearest neighbor interpolation.

[0069] (3) The no-displacement boundary condition is applied to the edge of the inner hole, and the free boundary condition is applied to the rectangular boundary of the main board. The finite element method is used to solve the thermal stress value on the non-structural grid point.

[0070] Based on the above steps, a dataset is constructed. The grid coordinates x, y and temperature in the dataset are then used as input to output thermal stress for model training.

[0071] Specifically, the unstructured grid thermal stress data set stores grid point coordinates, temperature, x-direction, y-direction, and xy-direction thermal stress.

[0072] Step 2: Construct a KD tree model optimized by Monte Carlo tree search, and adaptively decompose the computational domain into high stress gradient subdomains and low stress gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient. The flowchart is as follows: Figure 1 (a) The computational domain mentioned above refers to the area consisting of all points on the unstructured grid, which is the same as the area of ​​the satellite plate.

[0073] KD tree model optimized by Monte Carlo tree search, see Figure 1 As shown in (a), the KD tree treats all points as a tree structure, and each subdomain as a tree node, which contains the boundary position information between subdomains. The Monte Carlo tree search optimization algorithm is used to optimize the division of subdomains and the location of the boundaries between subdomains according to the optimization objective equation (2).

[0074] The point cloud coordinates are obtained when drawing an unstructured grid and refer to the coordinates of the unstructured grid nodes. The thermal stress gradient is obtained by interpolating the thermal stress with respect to the coordinates x and y using RBF (Radial Basis Function). The derivatives with respect to x and y reveal the relationship between the thermal stress gradients in the x and y directions. The gradient magnitude is then used as the basis for domain decomposition, using a KD tree model optimized with Monte Carlo tree search.

[0075] The optimization goal of the above domain decomposition is to minimize the sum of squares of stress gradients in each subdomain. It can be described as:

[0076]

[0077] Where, It consists of n non-intersecting rectangular subdomains, where n refers to the number of subdomains, and D(0) refers to the complete computational domain that has not been decomposed. i Indicates subdomain D (i) The sum of the squares of the differences between the gradient of each point in and the average gradient is as follows:

[0078]

[0079] Where M (i) Indicates D (i) The number of midpoints, Indicates subdomain D (i) The thermal stress gradient at the jth point in is obtained based on RBF calculation. Indicates subdomain D (i) Average value of the thermal stress gradient.

[0080] It should be noted that the above equation (2) is a complex combinatorial optimization problem. The solution of this problem is approximated by the Monte Carlo tree search algorithm. The solution of the problem refers to the result of regional decomposition. Subsequently, different prediction models will be used for sub-regions with different stress distribution characteristics to balance the prediction accuracy and calculation time.

[0081] In addition, each node of the Monte Carlo tree represents a KD tree and can take an action to move to its child nodes.

[0082] The action of the Monte Carlo Tree Search agent consists of three parts, including selecting the subdomain D * , choose a dimension from d * and using a hyperplane x d =b * Divide the selected subdomain D * .

[0083] The above is the simulation process of the Monte Carlo Tree Search algorithm, which is used to solve the steps in Equation 2. During a simulation, the Monte Carlo Tree Search algorithm repeatedly selects subdomains, dimensions, and hyperplanes based on the node value until the number of repetitions reaches a set value. The reward value is then calculated and used for the next simulation. When the number of simulations reaches the set value, the result is the domain decomposition result.

[0084] In order to evaluate the subdomain division results at the end of each agent simulation, based on the region D (0) TSS in and TSS in all subdomains i The sum of defines the final reward, which can be expressed as:

[0085]

[0086] The above formula (4) is used to calculate the reward value of the updated node in the back-propagation phase of the Monte Carlo tree search algorithm to guide the next simulation process.

[0087] Step 3: Build a high-precision physical field proxy model for multiple neural operators. Based on the stress distribution characteristics of the subdomains and the respective advantages of the two neural operators, Geo-FNO and NU-FNO are applied to the high stress gradient subdomain and the low stress gradient subdomain, respectively.

[0088] In high-gradient subdomains, Geo-FNO learns a deformable grid and solver operators. It uses a multilayer perceptron to transform the unstructured grid into a structured grid in latent space. It then uses a fast Fourier transform to learn the mapping from the temperature field to the thermal stress field. The algorithm uses a fully connected network to transform the unstructured grid into an adaptive grid and directly learns operators based on the unstructured data, avoiding errors caused by interpolation in high-stress gradient subdomains.

[0089] In the low-stress gradient subdomain, NU-FNO interpolates the point cloud in each subdomain into a uniform grid and uses grid technology based on fast Fourier transform to learn the solution operator. Specifically, NU-FNO interpolates the point cloud (unstructured data) of the low-stress gradient subdomain into structured data, and then uses fast Fourier transform to learn the mapping from temperature field to thermal stress field; compared with learning directly from unstructured data, it is faster and more accurate.

[0090] Therefore, the complementarity of the two neural operators better balances the efficiency and accuracy of thermal stress prediction. The flowchart of step 3 is as follows: Figure 1 As shown in (b), it can be broken down into five main steps:

[0091] Data preprocessing: First, in the high stress gradient subdomain In the example, the physical grid T is connected through a fully connected network. h,P Transformed into an adaptive grid T on the computational domain via a multilayer perceptron h,C, the physical grid refers to the unstructured grid used to create the dataset. In the low stress gradient region , input point cloud Interpolate to the input subdomain grid T l,C The input point cloud is specifically the point cloud of the unstructured grid thermal stress data set constructed in step 1. The point cloud (unstructured data) contains node coordinates, temperature, and stress. The output is: adaptive grid T h,C and subdomain grid T l,C The point cloud data (unstructured data) on the image contains node coordinates, temperature, and stress.

[0092] Data Dimensionality Upscaling: Adaptive Grid T h,C Temperature embedding in Transformed into hidden layer embedding through a fully connected network. In the subdomain grid T l,C Medium temperature embedding It is also converted into hidden layer embedding through a fully connected network, and the output is a high-dimensional latent variable in the latent space.

[0093] Information transmission: Grid T h,C and T l,C The hidden layer embedding in F is respectively transformed by geometric Fourier transform. a 、 and Fourier transform F, F -1 By passing L rounds, effective modeling and feature extraction of high-dimensional data can be achieved, thereby improving the expressive power of the model.

[0094] Data dimensionality reduction: hidden layer embeddings are transformed into computational grid T through a fully connected network h,C and T l,C Through this data dimensionality reduction, the key features in the hidden layer embedding are retained, redundant or noisy information is filtered out, and the downstream task (thermal stress prediction) is adapted to make the output of the data dimensionality reduction and the output vector of the data preprocessing the same size.

[0095] Data post-processing: Calculation grid T h,C and T l,C Deformed by multi-layer perceptron and linearly interpolated back to the physical grid T h,P and T l,P The final thermal stress field is obtained. See the attached Figure 1 Output thermal stress field in .

[0096] The example technical solution of this embodiment uses multiple neural operators to learn the stress characteristics of each subdomain. The specific process is:

[0097] (1) Divide the KD tree optimized by the Monte Carlo tree search algorithm into multiple non-overlapping rectangular subdomains.

[0098] (2) The average stress gradient of all points in the entire computational domain is used as the threshold to divide these subdomains into high stress gradient subdomains and low stress gradient subdomains.

[0099] (3) The GEO-FNO model is applied in the high stress gradient subdomain, and the NU-FNO model is applied in the low stress gradient subdomain to predict thermal stress.

[0100] Different prediction models are used for sub-regions with different stress distribution characteristics to balance prediction accuracy and calculation time.

[0101] Specifically, because the Fourier Neural Network Operator (FNO) is not adept at capturing information (high-frequency information) in regions with high stress gradients, and FNO (including NU-FNO) can only be applied to structured data, unstructured data must first be interpolated into structured data before FNO can be applied. Interpolation errors introduced in high-stress gradient subdomains exacerbate the poor learning performance and large prediction errors of FNO in these subdomains. GEO-FNO, however, uses a multilayer perceptron to transform the unstructured mesh into a structured mesh in high-stress gradient regions, thus avoiding interpolation errors.

[0102] In low-stress gradient regions, interpolation errors are small. The thermal stress prediction error based on the NU-FNO model is smaller than that obtained with the learned mesh deformation, and the interpolation calculation speed is faster. Therefore, using different prediction models in the two sub-regions can balance prediction accuracy and calculation time.

[0103] Example 2

[0104] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0105] Example 3

[0106] The purpose of this embodiment is to provide a computer-readable storage medium.

[0107] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0108] Example 4

[0109] The purpose of this embodiment is to provide a satellite thermal stress rapid prediction system based on domain decomposition, including:

[0110] The unstructured grid thermal stress dataset construction module is configured to: construct an unstructured grid thermal stress dataset consisting of temperature and stress data pairs, and use local grid refinement to obtain higher resolution high stress gradient areas;

[0111] A KD tree model building module is configured to: build a KD tree model optimized by Monte Carlo tree search, and adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient;

[0112] A high-precision physics proxy model construction module is configured to: construct a high-precision physics proxy model of multiple neural operators, apply Geo-FNO and NU-FNO in the high stress gradient subdomain and the low stress gradient subdomain respectively;

[0113] The thermal stress prediction module is configured as follows: in the high stress gradient subdomain, Geo-FNO is used to learn directly on unstructured data; in the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid and then learns the mapping based on fast Fourier transform to predict thermal stress under various thermal loads and complex geometries.

[0114] Example 5

[0115] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments.

[0116] The steps involved in the apparatus of the above embodiment correspond to those of the method embodiment 1. For detailed implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0117] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0118] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for rapid prediction of satellite thermal stress based on domain decomposition, characterized by: include: Construct an unstructured grid thermal stress dataset consisting of temperature and stress data pairs, and use local grid refinement to obtain higher resolution in high stress gradient areas; A KD tree model optimized by Monte Carlo tree search is constructed to adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient; Construct a high-precision physical field proxy model of multiple neural operators, and apply Geo-FNO and NU-FNO in the high stress gradient subdomain and low stress gradient subdomain respectively; In the high stress gradient subdomain, Geo-FNO is used to learn directly on unstructured data. In the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid and then learns a mapping based on fast Fourier transform to predict thermal stresses under various thermal loads and complex geometries.

2. The method for rapid prediction of satellite thermal stress based on domain decomposition according to claim 1, wherein: The process of constructing an unstructured grid thermal stress dataset consisting of temperature and stress data pairs is as follows: Input temperature field, which is a sample in Gaussian random field; When adjusting the mean and covariance of the Gaussian random field, the roughness of the temperature field changes; No-displacement boundary conditions are applied to the edges of the inner hole, and free boundary conditions are applied to the rectangular boundaries of the main plate; The thermal stress field is calculated using the finite element method.

3. The method for rapid prediction of satellite thermal stress based on domain decomposition according to claim 1, wherein: A KD tree model optimized by Monte Carlo tree search is constructed to adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient, including: The optimization goal of domain decomposition is to minimize the sum of squares of stress gradients in each subdomain; Each node of the Monte Carlo tree represents a KD tree and can take an action to move to its child node; The action of the Monte Carlo Tree Search agent consists of three parts, including selecting the subdomain D * , choose a dimension from d * and using a hyperplane x d =b * Divide the selected subdomain D * .

4. The method for rapid prediction of satellite thermal stress based on domain decomposition according to claim 1, wherein: For regional decomposition, the subdomain division results at the end of each agent simulation are evaluated based on the region D (0) TSS in and TSS in all subdomains i The sum defines the final reward.

5. The method for rapid prediction of satellite thermal stress based on domain decomposition according to claim 1, wherein: The optimization objective of domain decomposition is described as: Where, Consists of n non-intersecting rectangular subdomains, TSS i Indicates subdomain D (i) The sum of the squares of the differences between the gradient of each point in and the average gradient is as follows: Where M (i) Indicates D (i) The number of midpoints, Indicates subdomain D (i) The magnitude of the thermal stress gradient. Indicates subdomain D (i) average value of the thermal stress gradient; Equation (2) approximates the solution to this problem via the Monte Carlo tree search algorithm.

6. The method for rapid prediction of satellite thermal stress based on domain decomposition according to claim 1, wherein: A high-precision physical field proxy model of multiple neural operators is constructed. Geo-FNO and NU-FNO are applied in the high stress gradient subdomain and low stress gradient subdomain respectively. The detailed processing process is as follows: Data preprocessing: First, in the high stress gradient subdomain In the example, the physical grid T is connected through a fully connected network. h,P Deformed to an adaptive grid T on the computational domain h,C ; In the low stress gradient region , input point cloud Interpolate to the input subdomain grid T l,C ; Data Dimensionality Upscaling: Adaptive Grid T h,C Temperature embedding in Transformed into hidden layer embedding through a fully connected network, in the subdomain grid T l,C Medium temperature embedding It is also converted into hidden layer embedding through a fully connected network; Information transmission: Grid T h,C and T l,C The hidden layer embedding in F is respectively transformed by geometric Fourier transform. a 、F a -1 and Fourier transform F, F -1 Pass L wheel; Data dimensionality reduction: hidden layer embeddings are transformed into computational grid T through a fully connected network h,C and T l,C middle; Data post-processing: Calculation grid T h,C and T l,C Deform and interpolate back to the physical mesh T separately h,P and T l,P The final thermal stress field is obtained.

7. A satellite thermal stress rapid prediction system based on domain decomposition, characterized by: include: The unstructured grid thermal stress dataset construction module is configured to: construct an unstructured grid thermal stress dataset consisting of temperature and stress data pairs, and use local grid refinement to obtain higher resolution high stress gradient areas; A KD tree model building module is configured to: build a KD tree model optimized by Monte Carlo tree search, and adaptively decompose the domain into high-gradient subdomains and low-gradient subdomains according to the coordinates of the point cloud and the thermal stress gradient; A high-precision physics proxy model construction module is configured to: construct a high-precision physics proxy model of multiple neural operators, apply Geo-FNO and NU-FNO in the high stress gradient subdomain and the low stress gradient subdomain respectively; The thermal stress prediction module is configured as follows: in the high stress gradient subdomain, Geo-FNO is used to learn directly on unstructured data; in the low stress gradient subdomain, NU-FNO interpolates the point cloud into a uniform grid and then learns the mapping based on fast Fourier transform to predict thermal stress under various thermal loads and complex geometries.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method described in any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are performed.

Images