Online complex high-dimensional electromagnetic structure automatic design method, equipment and medium

Abstract

The invention discloses an automatic design method and device for an online complex high-dimensional electromagnetic structure and a medium. The method comprises the steps that an electromagnetic structure is modeled into a binary variable matrix, an optimization objective function is set, and an initial data set is constructed; training a proxy model by adopting the initial data set to obtain an initial proxy model; carrying out sensitivity analysis on the variables by utilizing a Sobol index; key points are extracted according to the contribution proportion, and optimization training is carried out on the proxy model; determining an initial solution and parameters of a simulated annealing algorithm, generating a neighborhood solution, calculating a target value and accepting strategy iterative optimization according to probability; performing simulation verification and online proxy model updating; key points are searched again, optimization is executed, and finally the sample with the maximum target value is screened out to serve as an optimization result. The method can effectively improve the optimization efficiency of the electromagnetic structure, reduces the calculation complexity and resource consumption, improves the optimization accuracy and reliability, and has a wide application prospect in the field of electromagnetic structure design.

Description

Technical Field

[0001] The present invention relates to the field of electromagnetic structure design, and in particular to an online complex high-dimensional electromagnetic structure automatic design method, equipment and medium. Background Art

[0002] With the widespread application of machine learning technology in the automated design of electromagnetic structures, how to improve the optimization efficiency in high-dimensional space has become an important research direction. Most existing technical solutions introduce proxy models (such as Gaussian processes, support vector machines or deep neural networks) to replace commercial simulation software with high computational costs to reduce the computational complexity of the optimization process, and directly combine heuristic algorithms to perform global searches in high-dimensional parameter spaces to find the optimal solution that meets multi-objective design requirements.

[0003] Although the above methods have improved the optimization efficiency to a certain extent, the existing technical solutions still have significant deficiencies in practical applications. The complexity of the high-dimensional design space makes it difficult to ensure the quality and representativeness of the samples, which makes it difficult for the proxy model to accurately fit the global characteristics of the objective function. In addition, the existing technical solutions usually directly optimize all input variables, the system has high computational complexity, and the optimization process wastes a lot of computing resources in redundant dimensions, which is inefficient and significantly limits its application effect in electromagnetic structure design.

[0004] Although the design of electromagnetic structures usually involves a high-dimensional design space, a large number of variables have a relatively weak impact on the target performance, while a few key variables play a decisive role in the optimization results. Therefore, a new technical solution is urgently needed that can effectively use the surrogate model to achieve effective dimensionality reduction of the high-dimensional design space without increasing the simulation cost, so as to improve the efficiency and overall performance of high-dimensional optimization. Summary of the invention

[0005] In order to solve at least one of the technical problems existing in the prior art to a certain extent, the object of the present invention is to provide an online complex high-dimensional electromagnetic structure automatic design method, device and medium based on sensitivity analysis.

[0006] The first technical solution adopted by the present invention is:

[0007] An online complex high-dimensional electromagnetic structure automatic design method comprises the following steps:

[0008] Model the electromagnetic structure as a binary variable matrix, set the optimization objective function, and construct the initial data set;

[0009] The proxy model is trained using the initial data set to obtain an initial proxy model;

[0010] The Sobol index is used to perform sensitivity analysis on matrix variables, a sample set is generated through Latin hypercube sampling, and the relevant variance and index are calculated based on the proxy model to quantify the contribution of the variables;

[0011] Extract key points based on contribution ratio and optimize the proxy model for training;

[0012] Simulated annealing optimization steps: construct a simulated annealing algorithm, based on the optimized proxy model, generate neighborhood solutions, calculate target values, and perform iterative optimization based on probability acceptance strategies;

[0013] After multiple rounds of iterations, some samples are simulated and verified, the data is merged and the proxy model is updated;

[0014] Recalculate the key points and perform optimization until the preset maximum number of optimization rounds is reached, and finally select the sample with the largest target value as the optimization result.

[0015] Furthermore, the electromagnetic structure is modeled as a binary variable matrix, an optimization objective function is set, and an initial data set is constructed, including:

[0016] The electromagnetic structure is modeled as an n*m dimensional binary variable matrix, where each element in the matrix represents a material unit;

[0017] Setting objective functions for quantifying electromagnetic structure performance;

[0018] Generate N by uniform random sampling sim The performance curve and target value are calculated by simulation software to form the initial data set D0.

[0019] Furthermore, the ResNet50 model is used as a proxy model;

[0020] The initial data set is used to train the proxy model to obtain the initial proxy model, including:

[0021] Preprocess the initial data set, use the preprocessed data set to train the proxy model, and obtain the initial proxy model θ0 is the parameter corresponding to the initial proxy model;

[0022] Among them, the loss function in the training process is defined as:

[0023]

[0024] In the formula, λ curve With λ target is the weight coefficient used to balance the curve loss and the target loss; n is the number of performance curves; is the model’s predicted value for the nth performance curve, S n(x) is the nth true performance curve of the electromagnetic structure calculated by simulation software; is the predicted value of the model for the target, and O(x) is the actual target value of the electromagnetic structure.

[0025] Furthermore, the Sobol index is used to perform sensitivity analysis on matrix variables, a sample set is generated by Latin hypercube sampling, and relevant variances and indices are calculated based on the proxy model to quantify the contribution ratio of variables, including:

[0026] The Sobol index is used to complete the sensitivity analysis of matrix variables, and Latin hypercube sampling is used to ensure that the values ​​of each variable are evenly distributed in the sample space; let the expected generated samples be N pre For each variable x i , each generates a random length of N pre A sequence of 0 and 1; the generated n*m N pre The length sequences are combined to obtain a uniformly distributed sample set X for sensitivity analysis. pre ;

[0027] Based on the initial proxy model Calculate the total variance Var(Y) and the input variable x i The contribution to the output variance Var(Y i );

[0028] According to the calculated Var(Y) and Var(Y i ), calculate the first-order Sobol index S i , which indicates the contribution of a single input variable to the output variance.

[0029] Furthermore, the initial proxy model Calculate the total variance Var(Y) and the input variable x i The contribution to the output variance Var(Y i ),include:

[0030] The contribution of each input variable to the output is measured by variance decomposition. Then the total variance decomposition formula of Y is:

[0031]

[0032] Where Y i is the variable x i The impact on the output variance, which measures the effect of changing x while keeping other variables fixed. i The impact of the value on the output; d is the total number of variables, Y ij Refers to variable xi With x j The impact of interactions between them on output;

[0033] Adopting an agent model For the sample set X pre Calculate the target value Y k , and use this to calculate the mean:

[0034]

[0035] Add weight value ω to the sample k , the new formula is defined as:

[0036]

[0037] Based on the obtained mean, calculate the total output variance:

[0038]

[0039] Calculate for each variable x i The conditional expectation of i When x is 0, filter out the high and low samples i For samples with 0, the conditional expectation is calculated according to the following formula:

[0040]

[0041] For a fixed x i For the case of 1, we can calculate x i The conditional expectation E(Y|x i =1), and further calculate the conditional expectation E(Y|x i ):

[0042] E(Y|x i )=p i E(Y|x i =0)+(1-p i )E(Y|x i =1)

[0043] In the formula, p i is the x in the sample i The probability of being 0; after completing the calculation of conditional expectation, the conditional expectation variance of each variable is calculated by the following formula:

[0044]

[0045] First-order Sobol index S i The calculation formula is:

[0046]

[0047] Where S i The larger the value of , the greater the value of variable x i The more significant the impact on the model output.

[0048] Furthermore, extracting key points according to contribution ratios and optimizing the proxy model includes:

[0049] The first-order Sobol index S of each input variable is calculated i Sort from largest to smallest and take the top 64 variables as key variables, denoted as x imp ={x i1 ,x i2 ,…,x i64}, and retrain the proxy model using the dataset D0; the input matrix of the trained proxy model is no longer composed of the entire space, but is reconstructed into an 8*8 matrix input using only these 64 variables. The trained new proxy model is recorded as

[0050] Furthermore, the simulated annealing optimization step comprises:

[0051] A1. In data set D0, except for the key variable x imp The other variables are denoted as x unimp ; Filter out the sample with the largest target from the simulated data set D0, which is x imp With x unimp Assigned as the starting solution of the simulated annealing algorithm. During the optimization process, x unimp It will always be consistent with the initial solution to effectively reduce the dimension of the optimization space, and its corresponding target value is recorded as O curebt ;

[0052] A2. To optimize the target, a simulated annealing algorithm is constructed and initial parameters are set;

[0053] A3, random pair x imp One or more variables in the y-axis are inverted to generate a neighborhood solution x' imp , and calculate the target value corresponding to the neighborhood solution based on the surrogate model If o new <o curent , then accept the newly generated neighborhood solution; otherwise, ac Accept, the expression is:

[0054]

[0055] Iterate and update the current solution and temperature. If this step accepts the new solution, then x imp =x' imp , Ocurent =O new ; Otherwise, keep the original solution; adjust the temperature using the preset temperature decay rate α:

[0056] T=αT

[0057] A4. Continue to iterate the process of step A3 until the temperature T is lower than the minimum temperature T min Or the number of iterations reaches the maximum number of iterations; after the iteration is completed, the final sample is used as the starting solution, and the simulated annealing optimization step is repeated again until the number of iterations reaches the preset value; among them, the samples with the performance ranking in the front preset position generated during the iteration process will be updated to the same sample set x top middle.

[0058] Furthermore, after the multiple rounds of iterations, simulation verification is performed on some samples, data is merged and the proxy model is updated, including:

[0059] For the sample set x top The Top-K samples in the dataset are simulated and verified, and then merged into the dataset D0. The newly obtained dataset is recorded as D1.

[0060] Use the dataset D1 to train the proxy model and get a new proxy model And update the weight value ω k .

[0061] The second technical solution adopted by the present invention is:

[0062] An electronic device comprises a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, the code set or the instruction set is loaded and executed by the processor to implement an online complex high-dimensional electromagnetic structure automatic design method as described above.

[0063] The third technical solution adopted by the present invention is:

[0064] A computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, wherein the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement an online complex high-dimensional electromagnetic structure automatic design method as described above.

[0065] The fourth technical solution adopted by the present invention is:

[0066] A computer program product or a computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above method.

[0067] The beneficial effects of the present invention are as follows: the present invention proposes a spatial dimensionality reduction scheme based on sensitivity analysis, which calculates the sensitivity of input variables by combining high and low confidence samples obtained by simulation and proxy model prediction to reduce the prediction and optimization space, improve prediction accuracy and reduce search costs, and use an online model update method to gradually improve the accuracy of sensitivity analysis, so that the algorithm can significantly reduce the calculation and time costs while maintaining high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the embodiments of the present invention or the drawings of related technical solutions in the prior art are introduced below. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0069] Figure 1 It is a flowchart of the steps of an online complex high-dimensional electromagnetic structure automatic design method in an embodiment of the present invention;

[0070] Figure 2 It is a schematic diagram of an online complex high-dimensional electromagnetic structure automatic design method based on sensitivity analysis in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0072] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0073] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0074] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0075] With the in-depth application of machine learning technology in the field of electromagnetic structure automation design, improving the efficiency of high-dimensional space optimization has become an important research direction. The existing design method introduces a proxy model to replace the time-consuming simulation software, and combines the heuristic algorithm to explore the high-dimensional design space, which makes it difficult for the predictor to fit the accurate global characteristics of the objective function using sparse data, and also makes the optimization waste a lot of computing resources in redundant dimensions. In order to overcome these limitations, the present invention proposes an online complex high-dimensional electromagnetic structure automatic design scheme based on sensitivity analysis, which can effectively use the proxy model to complete the effective high-dimensional design space dimensionality reduction without increasing the simulation cost, and significantly improve the efficiency and overall performance of high-dimensional optimization. Specifically, the process of the scheme of the present invention can be divided into seven steps: (a) problem definition and mathematical modeling. (b) training the proxy model. (c) sensitivity analysis of variables. (d) extracting key points and optimizing the proxy model. (e) simulated annealing optimization. (f) simulation verification and online proxy model update. (g) re-finding key points and performing optimization.

[0076] Example 1

[0077] like Figure 1 and Figure 2 As shown, this embodiment provides an online complex high-dimensional electromagnetic structure automatic design method based on sensitivity analysis, which can effectively improve the optimization efficiency of electromagnetic structure, reduce computational complexity and resource consumption, improve optimization accuracy and reliability, and has broad application prospects in the field of electromagnetic structure design. The method specifically includes the following steps:

[0078] S1. Model the electromagnetic structure as a binary variable matrix, set the optimization objective function, and construct the initial data set.

[0079] In some embodiments, step S1 specifically includes the following steps:

[0080] S1-1: In this step, the electromagnetic structure needs to be modeled as a binary variable matrix composed of 01 matrices. Each element in the matrix represents a material unit, 1 represents a metal unit, and 0 represents a non-metal unit. According to the size of the target structure and the size constraints of the minimum processing unit, it is modeled as an n*m dimensional matrix.

[0081] S1-2: Set the objective function used to quantify the performance of the electromagnetic structure. Common objectives include gain, bandwidth, etc. By using simulation software (such as CST STUDIO, HFSS, etc.), we can simulate each structure and calculate its corresponding performance curve S n (x). Where x represents the matrix corresponding to the structure, and n represents different curve types, such as S 21 Although electromagnetic structures have various performance curves, in the design process, we often only focus on certain performances in certain frequency bands, such as the maximum gain in the 5-6GHz frequency band. Therefore, we need to further calculate the target value corresponding to the structure. n (x). In order to simplify multi-objective optimization, we further calculate the total objective value O(x) to assist subsequent optimization tasks, which is defined as the objective value with the largest difference from the preset objective.

[0082] S1-3: Generate N by uniform random sampling sim The performance curve and target value of the matrix are calculated by simulation software to form the initial data set D0. At the same time, due to the high time cost of simulation evaluation, N sim It is often necessary to set it to a smaller value.

[0083] S2. Use the initial data set to train the proxy model to obtain an initial proxy model.

[0084] As an optional implementation, step S2 includes the following steps:

[0085] S2-1: In order to speed up the optimization process, this embodiment uses the ResNet50 model as a proxy model to replace the time-costly full-wave simulation. In this step, the input data needs to be preprocessed first, and the data set selects D0 generated in step S1-3. In order to make the model more generalizable, the input n*m dimensional matrix will be upsampled to 224*224 dimensions. In terms of output, since the performance curve of the electromagnetic structure is generally composed of 1001 dimensions, in order to enhance the model's understanding of complex high-dimensional data and improve the characterization ability, the model will simultaneously predict the performance curve and target value of the structure. The training of the model is constrained by the mean square error (MSE) loss, and the specific loss function is defined by the following formula:

[0086]

[0087] Among them, λ curve With λ target is the weight coefficient used to balance the curve loss and the target loss, n is the number of performance curves, is the model’s predicted value for the nth performance curve, is the model's predicted value for the target. Based on this step, we will get the initial proxy model θ0 is the parameter corresponding to the initial proxy model. It is worth noting that in the subsequent steps By default, only the target prediction value is output.

[0088] S3. Use the Sobol index to perform sensitivity analysis on matrix variables, generate sample sets through Latin hypercube sampling, and calculate relevant variances and indices based on the proxy model to quantify the contribution of variables.

[0089] As an optional implementation, step S3 includes the following steps:

[0090] S3-1: The embodiment of the present invention uses the Sobol index to complete the sensitivity analysis of matrix variables. In order to reduce the computational complexity and avoid the costly complete enumeration, Latin hypercube sampling can be used to ensure that the value of each variable is evenly distributed in the sample space. Suppose the expected sample to be generated is N pre (N pre >>N sim ), for each variable x i (a total of n*m variables), each generates a random length of N pre A sequence of 0 and 1. The generated n*m N pre By combining the length sequences, we can get the uniformly distributed sample set X for sensitivity analysis. pre .

[0091] S3-2: Measure the contribution of each input variable to the output through variance decomposition. Then the total variance decomposition formula of Y is:

[0092]

[0093] Among them, Y i refers to the variable x i The impact on the output variance, which measures the effect of changing x while keeping other variables fixed. i The impact of the value on the output, d is the total number of variables, Y ij Refers to variable x i With x j The key to calculating the first-order Sobol index is to calculate the total variance Var(Y) and the input variable x i The contribution to the output variance Var(Y i ). This step uses the proxy model obtained from S2 For the sample set X pre Calculate the target value Y k , and use this to calculate the mean:

[0094]

[0095] However, due to Y k It is predicted by the proxy model, which is a low-confidence sample, while the sample in the dataset D0 is a high-confidence sample verified by simulation, so it is necessary to add a weight ω to each item k To improve the accuracy of sensitivity analysis as much as possible, each high confidence sample corresponds to ω k All are 1, ω of low confidence samples k The value is determined by the prediction accuracy of the proxy model and the current optimization round. The prediction accuracy of the proxy model is determined by the Kendall's Tau, which will be referred to as KTau in the future. It is a non-parametric statistic that measures the correlation between two variables. It is usually used to evaluate the rank correlation of two variables. Its formula is:

[0096]

[0097] The set x is the sample label of the data set D0, that is, the real simulation performance target corresponding to each sample, and the set y is the proxy model output corresponding to each sample.

[0098] Assume that the total number of optimization rounds is M max round, the current round is M, the weight of low confidence samples is ω k for When the number of optimization rounds M in the initial stage is 1, ωk for At this point, the desired calculation is defined by the following new formula:

[0099]

[0100] Using the obtained mean, the total output variance can be further calculated:

[0101]

[0102] S3-3: To calculate Var(Y i ), first we need to calculate each variable x i The conditional expectation of . For a fixed x i When x is 0, filter out the high and low samples obtained in the previous step. i For samples with 0, the conditional expectation is calculated according to the following formula:

[0103]

[0104] The high and low confidence sample weights ω k Same as the previous step. Similarly, for a fixed x i For the case of 1, x can be calculated in the same way i The conditional expectation E(Y|x i =1), and further calculate the conditional expectation E(Y|x i ):

[0105] E(Y|x i )=p i E(Y|x i =0)+(1-p i )E(Y|x i =1)

[0106] where p i is the x in the sample i The probability of being 0, that is, the probability of x in the sample i The proportion of zero samples to the total samples. After completing the calculation of conditional expectation, the conditional expectation variance of each variable can be calculated by the following formula:

[0107]

[0108] S3-4: Using Var(Y) and Var(Y i ), which can be used to complete the first-order Sobol index S, a key parameter for measuring sensitivity i The index represents the contribution of a single input variable to the output variance, and its mathematical expression is:

[0109]

[0110] By calculating the first-order Sobol index S of each input variable i , we can quantify the independent contribution of each input to the model output. S i The larger the value of , the greater the value of variable x i The more significant the impact on the model output.

[0111] S4. Extract key points based on contribution ratio and optimize the agent model for training.

[0112] Exemplarily, the first-order Sobol index S of each input variable calculated in step S3 is i Sort from largest to smallest and take out the top 64 variables as x imp ={x i1 ,x i2 ,…,x i64}, use the dataset D0 to complete the proxy model training of step S2 again, but the difference is that the input matrix of the proxy model trained in this step is no longer composed of the entire space, but only reconstructed into an 8*8 matrix input with these 64 variables, which effectively reduces the difficulty of the proxy model in predicting complex high-dimensional space. The new proxy model obtained by training is recorded as

[0113] S5. Simulated annealing optimization step: Construct a simulated annealing algorithm based on the optimized proxy model, and perform iterative optimization by generating neighborhood solutions, calculating target values, and accepting strategies based on probability.

[0114] In one embodiment, step S5 specifically includes:

[0115] S5-1: Let the remaining variables be x unimp . Select the sample with the largest target from the simulated data set D0, which is x imp With x unimp Assigned as the starting solution of the simulated annealing algorithm. During the optimization process, x unimp It will always be consistent with the initial solution to effectively reduce the dimension of the optimization space, and its corresponding target value is recorded as O curent .

[0116] S5-2: To optimize the target, a simulated annealing algorithm is constructed. Its core mechanism is to gradually reduce the "energy" level of the system by introducing the "temperature" parameter to find the global optimal solution. In the initial stage, the parameters of the simulated annealing algorithm need to be set. The initial temperature needs to be set to a higher value such as T0=1000, the temperature decay rate α is set to a floating point number in the interval (0.8, 0.99), and the minimum temperature T minSet to 1, and the maximum number of iterations max_iter is set to 1000.

[0117] S5-3: Random pair x imp One or more variables in the y-axis are inverted to generate a neighborhood solution x' imp , and calculate the target value corresponding to the neighborhood solution based on the surrogate model If O new <O curent Then accept the newly generated neighborhood solution, otherwise accept it with probability P ac Accept, which is defined by the following formula:

[0118]

[0119] Iterate and update the current solution and temperature. If this step accepts the new solution, then x imp =x' imp , O curent =O new , otherwise keep the original solution. It is worth noting that the top 50 performance samples generated in the process will be continuously updated and stored in the sample set x top Then adjust the temperature using the preset temperature decay rate according to the following formula:

[0120] T=αT

[0121] S5-4: Continue to iterate the process of S5-3 until the temperature T is lower than the minimum temperature T min Or the number of iterations reaches the maximum number of iterations max_iter. After the iteration is completed, the final sample is used as the starting solution and step S5 is repeated again until a total of 10 iterations are completed. The top 50 samples generated in the process will be updated to the same sample set x top middle.

[0122] S6. After multiple rounds of iterations, some samples are simulated and verified, the data is merged and the proxy model is updated.

[0123] For example, in order to enable the proxy model to quickly adapt to environmental changes and gradually improve the prediction accuracy and reliability of sensitivity analysis, the present invention applies an online model update method in the optimization process, so that it can adjust the strategy according to new situations in a timely manner, enhance the dynamic adaptability of the model, and improve the optimization efficiency. top The Top-K samples in S2 are simulated and verified, and merged into the dataset D0. The newly obtained dataset is recorded as D1. The D1 data is used to re-complete the proxy model training in the S2 process to obtain a new proxy model And calculate the KTau value of the model to update the weight value ω of the low confidence sample kAt this point, the model input is also consistent with that described in step S2. When this entire process is completed, the optimization round M is increased by one.

[0124] S7, recalculate the key points and perform optimization until the preset maximum optimization round is reached, and finally select the sample with the largest target value as the optimization result.

[0125] Specifically, steps S3 to S6 are executed cyclically, and each cycle is performed using the new proxy model updated online and the data set with the new simulation data added, until the optimization round M reaches the preset maximum optimization round M. max , and finally select the sample with the largest target value from all the samples verified by simulation as the final optimization result.

[0126] In general, existing automatic design methods of electromagnetic structures based on machine learning usually use simple proxy models combined with traditional heuristic algorithms to search the entire design space, resulting in poor global optimization capabilities and low efficiency of the algorithm. When dealing with complex high-dimensional space design problems, this method is difficult to guarantee sample quality and representativeness, and the proxy model is difficult to accurately fit the global characteristics of the objective function. At the same time, the optimization process wastes a lot of computing resources on redundant dimensions and cannot quickly find the optimal solution area. In contrast, the present invention innovatively proposes a spatial dimensionality reduction scheme based on sensitivity analysis. By combining simulation and proxy model predictions to obtain high and low confidence samples, the sensitivity of the input variables is calculated to reduce the prediction and optimization space, improve prediction accuracy and reduce search costs, and use an online model update method to gradually improve the accuracy of sensitivity analysis, so that the algorithm can significantly reduce the computational and time costs while maintaining high accuracy.

[0127] In summary, the method of the present invention includes at least the following advantages and beneficial effects:

[0128] (1) Through the sensitivity analysis technology based on the surrogate model, the present invention can significantly improve the optimization efficiency in electromagnetic structure design tasks, especially in complex high-dimensional design tasks. The proposed surrogate model can further improve the data representation ability and make the reasoning results more accurate.

[0129] (2) The existing technology usually requires searching the entire huge design space. However, the present invention uses the prediction data of the proxy model and combines the sensitivity of the variables calculated by high and low confidence samples to achieve reliable spatial dimensionality reduction without adding additional simulation costs. This greatly reduces the difficulty of model prediction and optimization search, and also reduces the required computing resources and costs.

[0130] (3) The present invention uses an online model updating method to gradually improve the prediction accuracy of the proxy model for good sample areas during the optimization process, and continuously improves the sensitivity value, further improving the optimization ability of the algorithm.

[0131] (4) The present invention has a wide range of application adaptability. It is not only applicable to electromagnetic structure design problems, but can also be applied to other structural design problems, such as metasurface design, photon design, bearing design, etc., significantly improving the efficiency and performance of these tasks. Therefore, the scheme of the present invention has high technical advantages and wide practical application value.

[0132] Example 2

[0133] An embodiment of the present invention further provides an electronic device, the electronic device comprising a processor and a memory, the memory storing at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the at least one program, the code set or the instruction set being loaded and executed by the processor to implement the following Figure 1 An online complex high-dimensional electromagnetic structure automatic design method is shown.

[0134] It is understood that the memory may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory includes a non-transitory computer-readable storage medium. The memory may be used to store instructions, programs, codes, code sets, or instruction sets. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function, instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data created according to the use of the server, etc.

[0135] The processor may include one or more processing cores. The processor uses various interfaces and lines to connect the various parts of the entire server, and executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory, and calling data stored in the memory. Optionally, the processor can be implemented in at least one hardware form of digital signal processing (DSP), field programmable gate array (FPGA), and programmable logic array (PLA). The processor can integrate one or a combination of a central processing unit (CPU) and a modem. Among them, the CPU mainly processes the operating system and application programs; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor, but implemented separately through a chip.

[0136] Since the electronic device is an electronic device corresponding to an online complex high-dimensional electromagnetic structure automatic design method of an embodiment of the present invention, and the principle of solving the problem by the electronic device is similar to that of the method, the implementation of the electronic device can refer to the implementation process of the above-mentioned method embodiment, and the repeated parts will not be repeated.

[0137] Example 3

[0138] The embodiment of the present invention further provides a computer-readable storage medium, wherein the storage medium stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by a processor to implement the following Figure 1 An online complex high-dimensional electromagnetic structure automatic design method is shown.

[0139] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable rewritable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0140] Since the storage medium is the storage medium corresponding to an online complex high-dimensional electromagnetic structure automatic design method of an embodiment of the present invention, and the principle of solving the problem by the storage medium is similar to that of the method, the implementation of the storage medium can refer to the implementation process of the above-mentioned method embodiment, and the repeated parts will not be repeated.

[0141] Example 4

[0142] In some possible implementations, various aspects of the method of the embodiment of the present invention can also be implemented in the form of a program product, which includes a program code. When the program product is run on a computer device, the program code is used to enable the computer device to perform the steps of an online complex high-dimensional electromagnetic structure automatic design method according to various exemplary embodiments of the present application described above in this specification. Among them, the executable computer program code or "code" for executing each embodiment can be written in a high-level programming language such as C, C++, C#, Smalltalk, Java, JavaScript, Visual Basic, structured query language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0143] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0144] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0145] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable ordinary technicians in the field to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims

1. An online complex high-dimensional electromagnetic structure automatic design method, characterized in that: The following steps are involved: Model the electromagnetic structure as a binary variable matrix, set the optimization objective function, and construct the initial data set; The proxy model is trained using the initial data set to obtain an initial proxy model; The Sobol index is used to perform sensitivity analysis on matrix variables, a sample set is generated through Latin hypercube sampling, and the relevant variance and index are calculated based on the proxy model to quantify the contribution of the variables; Extract key points based on contribution ratio and optimize the proxy model for training; Simulated annealing optimization steps: construct a simulated annealing algorithm, based on the optimized proxy model, generate neighborhood solutions, calculate target values, and perform iterative optimization based on probability acceptance strategies; After multiple rounds of iterations, some samples are simulated and verified, the data is merged and the proxy model is updated; Recalculate the key points and perform optimization until the preset maximum number of optimization rounds is reached, and finally select the sample with the largest target value as the optimization result.

2. The online complex high-dimensional electromagnetic structure automatic design method according to claim 1 is characterized in that: The electromagnetic structure is modeled as a binary variable matrix, an optimization objective function is set, and an initial data set is constructed, including: The electromagnetic structure is modeled as an n*m dimensional binary variable matrix, where each element in the matrix represents a material unit; Setting objective functions for quantifying electromagnetic structure performance; Generate n by uniform random sampling sim The performance curve and target value are calculated by simulation software to form the initial data set D0.

3. The online complex high-dimensional electromagnetic structure automatic design method according to claim 1 is characterized in that: The initial data set is used to train the proxy model to obtain the initial proxy model, including: Preprocess the initial data set, use the preprocessed data set to train the proxy model, and obtain the initial proxy model θ0 is the parameter corresponding to the initial proxy model; Among them, the loss function in the training process is defined as: In the formula, λ curve With λ target is the weight coefficient used to balance the curve loss and the target loss; n is the number of performance curves; is the model’s predicted value for the nth performance curve, S n (x) is the nth true performance curve of the electromagnetic structure calculated by simulation software; is the predicted value of the model for the target, and O(x) is the actual target value of the electromagnetic structure.

4. The method for automatic design of an online complex high-dimensional electromagnetic structure according to claim 1, characterized in that: The Sobol index is used to perform sensitivity analysis on matrix variables, a sample set is generated through Latin hypercube sampling, and relevant variances and indices are calculated based on the proxy model to quantify the contribution ratio of variables, including: The Sobol index is used to complete the sensitivity analysis of matrix variables, and Latin hypercube sampling is used to ensure that the values ​​of each variable are evenly distributed in the sample space; let the expected generated samples be N pre For each variable x i , each generates a random length of N pre The sequence of n*m N pre The length sequences are combined to obtain a uniformly distributed sample set X for sensitivity analysis. pre ; Based on the initial proxy model Calculate the total variance Var(Y) and the input variable x i The contribution to the output variance Var(Y i ); According to the calculated Var(Y) and Var(Y i ), calculate the first-order Sobol index S i , which indicates the contribution of a single input variable to the output variance.

5. The online complex high-dimensional electromagnetic structure automatic design method according to claim 4 is characterized in that: Based on the initial proxy model Calculate the total variance Var(Y) and the input variable x i The contribution to the output variance Var(Y i ),include: Adopting an agent model For the sample set X pre Calculate the target value Y k , and use this to calculate the mean: Add weight value ω to the sample k , the new formula is defined as: Based on the obtained mean, calculate the total output variance: Calculate for each variable x i The conditional expectation of i When x is 0, filter out the high and low samples i For samples with 0, the conditional expectation is calculated according to the following formula: For a fixed x i For the case of 1, we can calculate x i The conditional expectation E(Y|x i =1), and further calculate the conditional expectation E(Y|x i ): E(Y|x i )=p i E(Y|x i (0)+(1-p) i )E(Y|x i =1) In the formula, p i is the x in the sample i The probability of being 0; after completing the calculation of conditional expectation, the conditional expectation variance of each variable is calculated by the following formula: First-order Sobol index S i The calculation formula is: Where S i The larger the value of , the greater the value of variable x i The more significant the impact on the model output.

6. The online complex high-dimensional electromagnetic structure automatic design method according to claim 1, characterized in that: The key points are extracted according to the contribution ratio, and the proxy model is optimized and trained, including: The first-order Sobol index S of each input variable is calculated i Sort from largest to smallest and take the top 64 variables as key variables, denoted as x imp ={x i1 ,x i2 ,…,x i64 }, and retrain the proxy model using the dataset D0; the input matrix of the trained proxy model is no longer composed of the entire space, but is reconstructed into an 8*8 matrix input using only these 64 variables. The trained new proxy model is recorded as 7. The online complex high-dimensional electromagnetic structure automatic design method according to claim 1 is characterized in that: The simulated annealing optimization step comprises: A1. In data set D0, except for the key variable x imp The other variables are denoted as x unimp ; Filter out the sample with the largest target from the simulated data set D0, which is x imp With x unimp Assigned as the starting solution of the simulated annealing algorithm. During the optimization process, x unimp It will always be consistent with the initial solution to effectively reduce the dimension of the optimization space, and its corresponding target value is recorded as O curent ; A2. To optimize the target, a simulated annealing algorithm is constructed and initial parameters are set; A3, random pair x imp One or more variables in the y-axis are inverted to generate a neighborhood solution x' imp , and calculate the target value corresponding to the neighborhood solution based on the surrogate model If O new <O curent , then accept the newly generated neighborhood solution; otherwise, ac Accept, the expression is: Iterate and update the current solution and temperature. If this step accepts the new solution, then x imp =x' imp , O curent =O new ; Otherwise, keep the original solution; adjust the temperature using the preset temperature decay rate α: T=αT A4. Continue to iterate the process of step A3 until the temperature T is lower than the minimum temperature T min Or the number of iterations reaches the maximum number of iterations; after the iteration is completed, the final sample is used as the starting solution, and the simulated annealing optimization step is repeated again until the number of iterations reaches the preset value; among them, the samples with the performance ranking in the front preset position generated during the iteration process will be updated to the same sample set x top middle.

8. The online complex high-dimensional electromagnetic structure automatic design method according to claim 1, characterized in that: After the multiple rounds of iterations, some samples are simulated and verified, data is merged and the proxy model is updated, including: For the sample set x top The Top-K samples in the dataset are simulated and verified, and then merged into the dataset D0. The newly obtained dataset is recorded as D1. Use the dataset D1 to train the proxy model and get a new proxy model And update the weight value ω k .

9. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the method according to any one of claims 1 to 8.

Images