Adapter Optimization Design Method Based on Parameter Optimization Neural Network

Abstract

The present invention proposes an adapter optimization design method based on a parameter optimization neural network, which belongs to the field of adapter design, including: S1, using Latin hypercube sampling method to generate adapter configuration samples, obtaining strain field data through a finite element model, and using the adapter configuration sample input parameters and strain field distribution as a training set to train an uncertainty quantification model; S2, using the trained uncertainty quantification model as an approximate proxy model to calculate the objective function; S3, constructing a parameter optimization neural network, and using the training set and the trained uncertainty quantification model to train the parameter optimization neural network to obtain the trained parameter optimization neural network, and the output of the trained parameter optimization neural network is the optimal design parameters of the adapter. This application uses the trained uncertainty quantification model to perform performance evaluation and parameter optimization on the parameter optimization neural network, and determines the optimal weight parameters, and then determines the optimal design parameters of the adapter, thereby improving the efficiency and accuracy of the adapter optimization design.

Description

Technical Field

[0001] The present invention relates to the technical field of adapter design, and in particular to an adapter optimization design method based on parameter optimization neural network. Background Art

[0002] Box launch is a common method of missile launch, in which the adapter is a key component connecting the missile and the launch box, which plays a role in fixing, shock absorbing and radially supporting the missile, and also plays a role in guiding the missile launch. In order to improve the storage safety and launch performance of the missile, it is necessary to optimize the structure of the adapter.

[0003] In the prior art, deep learning-based metamaterial optimization design methods usually use a neural network model (DNN) as an objective function evaluator, and combine it with intelligent optimization algorithms such as genetic algorithms to optimize the design variables. This method requires evaluating the objective function values ​​of multiple different individuals in the initialization phase to maintain the diversity of the population, and performing a large number of objective function calculations in each iteration. Since the evaluation of support performance requires geometric modeling and coordinate discretization, it still requires a certain amount of computing time, so the overall optimization process is relatively time-consuming.

[0004] Therefore, finding a method that can both reduce computational overhead, improve adapter design modeling efficiency, and improve evaluation support performance is a technical problem that needs to be urgently solved by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention proposes an adapter optimization design method based on a parameter optimization neural network, which can use the trained uncertainty quantification model to perform performance evaluation and parameter optimization on the parameter optimization neural network, and determine the optimal weight parameters, and then determine the optimal design parameters of the adapter, thereby improving the efficiency and accuracy of the adapter optimization design.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides an adapter optimization design method based on parameter optimization neural network, comprising the following steps:

[0007] S1. Use Latin hypercube sampling method to generate adapter configuration samples, obtain strain field data through finite element model, and use adapter configuration sample input parameters and strain field distribution as training set to train uncertainty quantification model;

[0008] S2, using the trained uncertainty quantification model as an approximate proxy model to calculate the objective function;

[0009] S3. Construct a parameter optimization neural network, and use the training set and the trained uncertainty quantification model to train the parameter optimization neural network to obtain the trained parameter optimization neural network, wherein the output of the trained parameter optimization neural network is the optimal design parameter of the adapter.

[0010] On the basis of the above technical solution, preferably, the training process of the parameter optimization neural network specifically includes:

[0011] Input the training set into the parameter optimization neural network and initialize the parameter optimization neural network;

[0012] Calculate the support performance of each adapter configuration sample point in the training set, form a support performance sequence, determine the maximum value in the support performance sequence, and use the maximum value as the current design parameter of the adapter;

[0013] Back propagation is used to train the randomly initialized parameter optimization neural network so that the output value of the parameter optimization neural network is close to the design variable that supports the maximum performance, thereby obtaining the a priori parameter initialization parameters;

[0014] The parameter optimization neural network is iteratively optimized using the trained uncertainty quantification model and the priori parameter initialization parameters to obtain the optimal model parameters;

[0015] The optimal design parameters of the adapter are determined according to the optimal model parameters, and the output of the uncertainty quantification model is the optimal design parameters of the adapter.

[0016] On the basis of the above technical solution, preferably, the pre-training of the parameter optimization neural network specifically includes:

[0017] The AdamW optimizer is used to optimize the parameters according to the loss function of the parameter optimization neural network and the weight parameters of the current parameter optimization neural model to obtain the prior parameter initialization parameters.

[0018] On the basis of the above technical solution, preferably, the iterative optimization process specifically includes:

[0019] Step 1: Determine the current design parameters of the adapter based on the a priori parameter initialization parameters;

[0020] Step 2: Perform geometric deformation and mesh generation according to the current design parameters of the adapter to obtain sampling points;

[0021] Step 3: Use the trained uncertainty quantification model to evaluate the support performance according to the current design parameters of the sampling point and the adapter, and obtain the loss function and objective function of the current sampling point;

[0022] Step 4: Compare the offspring objective function with the optimal objective function and update it. Use the ADAMS optimizer to update the parameters of the parameter optimization neural network to obtain a new generation of parameters.

[0023] Step 5: Repeat steps 1 to 4 until the iteration stop condition is reached, and save the optimized weight parameters of the parameter optimization neural network after iterative optimization.

[0024] Based on the above technical solution, preferably, the trained uncertainty quantification model is used to perform support performance evaluation according to the current design parameters of the sampling points and the adapter, specifically including:

[0025] Input the current design parameters of the sampling points and the adapter into the trained uncertainty quantification model and calculate the confidence interval of the sampling points;

[0026] The objective function is determined based on the confidence interval, and the minimum value of the objective function is determined.

[0027] Based on the above technical solution, preferably, the calculation formula of the confidence interval is:

[0028]

[0029] Among them, x represents the sampling point, G CNFA-UQ represents the uncertainty quantification model after training, M represents the index of the confidence interval, and L k represents the design parameters of the adapter at the kth iteration, Represents the current design parameters of the adapter.

[0030] On the basis of the above technical solution, preferably, the calculation formula of the objective function is:

[0031]

[0032] Among them, J k represents the support performance, sup(CI) represents the upper bound of the confidence interval CI, N k Indicates the number of sampling points at the kth iteration.

[0033] More preferably, the training set includes design variables, loads and strain field data of adapter configuration sample points.

[0034] On the basis of the above technical solution, preferably, the adapter is composed of a honeycomb unit cell structure, and the number of design variables of the adapter is 6.

[0035] The adapter optimization design method of the present invention has the following beneficial effects compared with the prior art:

[0036] (1) The adapter configuration samples were effectively generated by the Latin hypercube sampling method, and the corresponding strain field data were obtained by using the finite element model. The adapter configuration samples and strain field data were used as training sets to train the uncertainty quantification model. The trained uncertainty quantification model was used to evaluate the performance and optimize the parameters of the parameter optimization neural network, and the optimal weight parameters were determined, thereby determining the optimal design parameters of the adapter, thereby improving the efficiency and accuracy of the adapter optimization design.

[0037] (2) The adapter configuration sample points and support information in the training set are used as the training data of the parameter optimization neural network. The trained uncertainty quantification model is used to evaluate the performance of the current design parameters of the adapter in each iterative optimization, so that the weight parameters of the parameter optimization neural network gradually approach the optimal weight parameters, effectively avoiding the optimization distortion caused by uncertainty and improving the reliability and robustness of the adapter design optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 The basic configuration of the adapter of the adapter optimization design method of the present invention, wherein (a) is a three-dimensional view and (b) is a front view;

[0040] Figure 2 A flow chart of the adapter optimization design method of the present invention;

[0041] Figure 3 A flowchart of the adapter optimization design method of the present invention;

[0042] Figure 4 The adapter automated finite element simulation process of the adapter optimization design method of the present invention, wherein (a) is the geometric configuration, (b) is the meshing, (c) is the load and boundary conditions, and (d) is the calculation solution;

[0043] Figure 5 A structural diagram of the adapter optimization design method of the present invention;

[0044] Figure 6 The objective function change process of the optimization process under the PONN, GA, and DE algorithms of the adapter optimization design method of the present invention;

[0045] Figure 7Deterministic and uncertain optimization processes under PONN, GA, and DE algorithms of the adapter optimization design method of the present invention;

[0046] Figure 8 The invention discloses an optimization process of different uncertainty quantification models under PONN, GA and DE algorithms of the parameter optimization neural network of the adapter optimization design method. DETAILED DESCRIPTION

[0047] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, in the embodiment provided by the present invention, the size of the adapter is 150mm×80mm×60mm, the material is polyurethane with superelastic mechanical properties, and the adapter is composed of a series of honeycomb unit cell structures, so the geometric shape of the adapter is controlled by the unit cell structure, and the geometric shape of the unit cell structure is determined by L 1 ~L 6 There are 6 design parameters in total; the design variable of the adapter is the design parameter L of the unit cell structure 1 ~L 6 , the number of design variables is 6.

[0049] like Figure 2 and Figure 3 As shown, the present invention provides an adapter optimization design method based on parameter optimization neural network, comprising the following steps:

[0050] S1. The adapter configuration samples are generated by Latin hypercube sampling method, strain field data are obtained through finite element model, and the input parameters and strain field distribution of the adapter configuration samples are used as training sets to train the uncertainty quantification model (CNFA-UQ model). The training set includes the design variables, loads and strain field data of the adapter configuration sample points.

[0051] Specifically, step S1 is described with a specific embodiment:

[0052] By adopting the Latin hypercube sampling method, 183 adapter configurations were generated, 20 loads of different sizes were applied to each configuration, and the total number of samples was 3660;

[0053] The adapter configuration is batch simulated according to the load through the finite element model, and the strain field data of the adapter at each adapter configuration sample point is obtained. The input parameters and strain field distribution of the adapter configuration sample are used as the training set to train the uncertainty quantification model (CNFA-UQ model).

[0054] In one embodiment of the present application, ANSYS Mechanical software is used to establish a finite element model, and the APDL command stream is used to perform automated meshing, boundary and load condition setting, and solving.

[0055] like Figure 4 As shown, MATLAB drives SOLIDWORKS to perform parametric modeling on the basic configuration to obtain the corresponding geometric body; then MATLAB drives ANSYS Mechanical to perform a series of operations such as meshing, setting boundary conditions and loads, and calculating and solving to obtain the final simulation results.

[0056] By using the adapter configuration sample input parameters and strain field distribution as the training set to train the uncertainty quantification model, the uncertainty characteristics of the adapter under various configurations and loading conditions can be accurately predicted, greatly improving the efficiency of adapter design.

[0057] S2. Use the trained uncertainty quantification model as an approximate proxy model to calculate the objective function.

[0058] It can be understood that the trained uncertainty quantification model calculates the strain field distribution and uncertainty distribution under each design variable, and evaluates the support performance of the objective function evaluator through the predicted confidence upper bound. Among them, the predicted confidence upper bound refers to a method of uncertainty quantification of the prediction result, which represents the upper limit of the prediction result under the confidence level. That is, the CNFA-UQ model learns the conditional probability distribution of the strain field during the training process, and uses the conditional probability distribution to calculate the predicted value of the strain field under the design variable and the corresponding confidence interval.

[0059] Specifically, 183 sample points of Latin super-mean square sampling are used as the data set. After geometric deformation and network partitioning, the trained CNFA-UQ model is used as the objective function evaluator. In order to prevent the uncertainty of the CNFA-UQ model from causing deviations in the optimization results, the strain field distribution and uncertainty distribution under each design variable are calculated, and the support performance is calculated with the predicted confidence upper bound. When the sample point is the optimal point obtained by the optimization algorithm, its actual support performance is greater than the predicted support performance, so the obtained result is better.

[0060] The strain field distribution and uncertainty distribution under the design variables are predicted by the uncertainty quantification model, the support performance of the adapter design scheme is accurately evaluated, and the predicted confidence upper bound is used as the optimization objective function to effectively avoid the deviation of the optimization results caused by the uncertainty of the uncertainty quantification model itself, thereby enhancing the credibility of the optimal solution of the parameter optimization neural network.

[0061] Specifically, step S2 is described with a specific embodiment:

[0062] Calculating the support performance of 183 simulation data, the maximum support performance is 321.06, and the corresponding design parameters are [7.92, 4.64, 3.16, 17.91, 9.24, 3.11], and the minimum support performance is 72.56, and the corresponding design parameters are [5.53, 2.86, 2.91, 18.92, 14.54, 2.42].

[0063] In one embodiment of the present application, the uncertainty quantification model (CNFA-UQ model) is used to predict the initial design parameters of the adapter. The predicted strain field distribution is almost consistent with the actual strain field distribution, and the average prediction error is 5.87%. The overall prediction error is small, and the positions with large prediction errors mainly appear at the end face connection and the upper and lower symmetrical center lines. The corresponding uncertainty quantification results are similar to the error distribution. The points with large errors in the error graph also have large corresponding uncertainties. At this time, the support performance is 306.53, and the support performance without considering the uncertainty is 317.80. The support performance obtained when considering the uncertainty is smaller.

[0064] S3, constructing a parameter optimization neural network, and using the training set and the trained uncertainty quantification model to train the parameter optimization neural network, to obtain a trained parameter optimization neural network, wherein the output of the trained parameter optimization neural network is the optimal design parameter of the adapter; wherein the parameter optimization neural network includes a feedforward neural network, such as Figure 5 shown.

[0065] It can be understood that the weight parameters of the parameter-optimized neural network represent the design parameters of the adapter.

[0066] Specifically, the training process of the parameter optimization neural network specifically includes:

[0067] Input the training set into the parameter optimization neural network and initialize the parameter optimization neural network;

[0068] Calculate the support performance of each adapter configuration sample point in the training set, form a support performance sequence, determine the maximum value in the support performance sequence, and use the maximum value as the current design parameter of the adapter;

[0069] Back propagation is used to train the randomly initialized parameter optimization neural network so that the output value of the parameter optimization neural network is close to the design variable that supports the maximum performance, thereby obtaining the a priori parameter initialization parameters;

[0070] The parameter optimization neural network is iteratively optimized using the trained uncertainty quantification model and the priori parameter initialization parameters to obtain the optimal model parameters;

[0071] The optimal design parameters of the adapter are determined according to the optimal model parameters, and the output of the uncertainty quantification model is the optimal design parameters of the adapter.

[0072] The adapter configuration sample points and support information in the training set are used as training data for the parameter optimization neural network, and these training data are used to enable the parameter optimization neural network to learn the complex mapping relationship between design parameters and support performance, and combined with the trained uncertainty quantification model for optimization, the robustness and credibility of the parameter optimization neural network are improved, thereby improving the efficiency of adapter design.

[0073] Furthermore, the pre-training of the parameter optimization neural network specifically includes:

[0074] The AdamW optimizer is used to optimize the parameters according to the loss function of the parameter optimization neural network and the weight parameters of the current parameter optimization neural model to obtain the prior parameter initialization parameters.

[0075] In one embodiment of the present application, the loss function of the parameter optimization neural network model is:

[0076]

[0077] in, It represents the weight parameters of the parameter optimization neural network when the design variables are 1, 2, 3, 4, 5, and 6, and L' represents the current design parameters of the adapter.

[0078] Furthermore, the iterative optimization process specifically includes:

[0079] Step 1: Determine the current design parameters of the adapter based on the a priori parameter initialization parameters;

[0080] Step 2: Perform geometric deformation and mesh generation according to the current design parameters of the adapter to obtain sampling points;

[0081] Step 3: Use the trained uncertainty quantification model to evaluate the support performance according to the current design parameters of the sampling point and the adapter, and obtain the loss function and objective function of the current sampling point;

[0082] Step 4: Compare the offspring objective function with the optimal objective function and update it. Use the ADAMS optimizer to update the parameters of the parameter optimization neural network to obtain a new generation of parameters.

[0083] Step 5: Repeat steps 1 to 4 until the iteration stop condition is reached, and save the optimized weight parameters of the parameter optimization neural network after iterative optimization.

[0084] By using the trained uncertainty quantification model to perform performance evaluation on the current design parameters of the adapter in each iterative optimization, the weight parameters of the parameter optimization neural network gradually approach the optimal weight parameters, that is, the optimal design parameters of the adapter, thereby improving the optimization efficiency of the adapter design, and at the same time effectively avoiding the optimization distortion caused by uncertainty, and improving the reliability and robustness of the adapter design optimization. In one embodiment of the present application, the iterative stop condition is that the support performance of the current parameter optimization neural network is the optimal support performance.

[0085] Furthermore, the trained uncertainty quantification model is used to support performance evaluation based on the sampling points and the current design parameters of the adapter, including:

[0086] Input the current design parameters of the sampling points and the adapter into the trained uncertainty quantification model and calculate the confidence interval of the sampling points;

[0087] The objective function is determined based on the confidence interval, and the minimum value of the objective function is determined.

[0088] Specifically, the calculation formula for the confidence interval is:

[0089]

[0090] Among them, x represents the sampling point, G CNFA-UQ represents the uncertainty quantification model after training, M represents the index of the confidence interval, and L k represents the design parameters of the adapter at the kth iteration, Represents the current design parameters of the adapter.

[0091] In one embodiment of the present application, the CNFA-UQ model outputs 20 confidence intervals, that is, the range of M is 1 to 20, and the confidence interval is:

[0092]

[0093] Specifically, the calculation formula of the objective function is:

[0094]

[0095] Among them, J krepresents the support performance, sup(CI) represents the upper bound of the confidence interval CI, N k Indicates the number of sampling points at the kth iteration.

[0096] This application effectively generates adapter configuration samples through the Latin hypercube sampling method, and obtains the corresponding strain field data using the finite element model. The adapter configuration samples and strain field data are used as training sets to train the uncertainty quantification model. The trained uncertainty quantification model is used to perform performance evaluation and parameter optimization on the parameter optimization neural network, and the optimal weight parameters are determined, thereby determining the optimal design parameters of the adapter, thereby improving the efficiency and accuracy of the adapter optimization design.

[0097] In one embodiment of the present application, the design variable dimension of the adapter is 6, the number of input and output neurons of the parameter optimization neural network (PONN) is 6, the number of hidden layers is 3, and the number of hidden layer neurons is 100. The input is an integer index vector from 1 to 6, that is, Represents the index of the corresponding design variable, and the output is a 6-dimensional floating-point vector Represents the value of the design variable.

[0098] In one embodiment of the present application, in the initialization stage of the parameter optimization neural network, the maximum number of iterations is set to 2000, the cosine annealing learning rate decay strategy is adopted, the initial learning rate is 0.01, the minimum learning rate is 0.0001, the cosine annealing period is 2000, and the maximum number of early stopping mechanisms is 100.

[0099] In one embodiment of the present application, the maximum number of iterations of iterative optimization is set to N. optim , after N optim After iterations, the optimal model parameters are obtained, and the output of the model is the optimal design parameters. Due to the randomness of the AdamW algorithm, each iteration step is not always in the direction of the loss function, so the support performance is not improved at each step. Here, the principle of the early stopping mechanism is adopted. When the support performance is improved, the current model weight file is saved, and the optimal support performance is set as the current support performance.

[0100] Preferably, in the initialization stage of the parameter optimization neural network model, the parameter optimization neural network model ended training due to the early stopping mechanism after 437 iterations, and the total training time was 1.87s. The loss function value decreased from the initial 2.02 to 3.79×10-14, and the output value of the parameter optimization neural network was [7.92, 4.64, 3.16, 17.91, 9.24, 3.11], which was equal to the design parameters of the maximum support performance, proving that the proposed parameter initialization method can effectively initialize the parameters of the parameter optimization neural network, so that the design parameters represented by the parameter optimization neural network are the current optimal design parameters of the adapter.

[0101] In one embodiment of the present application, during the iterative optimization process of the parameter optimization neural network model, the weight parameters of the parameter optimization neural network are set to a trainable mode, the parameters of the uncertainty quantification model (CNFA-UQ model) are all frozen, the output of the parameter optimization neural network is divided according to the geometric deformation, the uncertainty quantification model (CNFA-UQ model) is used to quantify the uncertainty of the IFE-ONet model and predict the strain field under different configurations. During the iterative optimization process, the weight parameters of PONN are the only trainable parameters. Since the optimization goal here is to maximize the support performance as much as possible, and J% is always positive, the maximum number of iterations is set to 100, the initial learning rate of cosine annealing is 0.01, the minimum learning rate is 0.001, and the maximum number of early stopping mechanisms is 20. The weight parameters of PONN are updated using the AdamW optimizer, and the optimal support performance of PONN starts from 305.64. Since the early stopping mechanism converges after 160 iterations, the final support performance is 362.08.

[0102] In one embodiment of the present application, the optimization results of the parameter optimization neural network are compared with the optimization results of the optimization method based on the intelligent optimization algorithm. The comparison results are shown in Table 1, where the optimization method based on the intelligent optimization algorithm includes genetic algorithm (Generic algorithm, GA) and differential evolution algorithm (Differential evolution, DE).

[0103] Table 1

[0104]

[0105] As shown in Table 1, the parameter optimization neural network of this application improves the support performance of the adapter by 4.10%, achieving good optimization results. Figure 6 As shown in the figure, the number of objective function evaluations is only one fifth of that in achieving similar optimization results.

[0106] In one embodiment of the present application, the optimization results of the parameter optimization neural network model using the uncertain quantitative model prediction are compared with the optimization results of the information fusion feature equalization operator model prediction value. The optimization process is as follows: Figure 6 The comparison results are shown in Table 2.

[0107] Table 2

[0108]

[0109] As can be seen from Table 2, the use of confidence upper bounds to evaluate support performance has achieved positive optimization results. Taking into account the impact of model prediction errors, the errors can be corrected to obtain more robust optimization results. The optimal support performance has increased by an average of 5.07%, while the deterministic optimization strategies that directly use model prediction results have achieved failure solutions, and the optimal support performance has decreased by an average of 7%, proving the effectiveness of the uncertainty optimization strategy of this application.

[0110] In one embodiment of the present application, the uncertainty quantification model used in the present application is analyzed with the MVDER, CQR and VB models. Figure 8 The comparison results are shown in Table 3.

[0111] Table 3

[0112]

[0113]

[0114] It can be seen from Table 3 that the results of uncertainty optimization using the CNFA-UQ of this application are better than those of the MVDER, CQR and VB models, which proves that the CNFA-UQ of this application has more accurate uncertainty quantification results.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An adapter optimization design method based on parameter optimization neural network, characterized in that: The following steps are involved: S1. Use Latin hypercube sampling method to generate adapter configuration samples, obtain strain field data through finite element model, and use adapter configuration sample input parameters and strain field distribution as training set to train uncertainty quantification model; S2, using the trained uncertainty quantification model as an approximate proxy model to calculate the objective function; S3, constructing a parameter optimization neural network, and using the training set and the trained uncertainty quantification model to train the parameter optimization neural network to obtain a trained parameter optimization neural network, wherein the output of the trained parameter optimization neural network is the optimal design parameter of the adapter; The training process of the parameter optimization neural network specifically includes: Input the training set into the parameter optimization neural network and initialize the parameter optimization neural network Calculate the support performance of each adapter configuration sample point in the training set, form a support performance sequence, determine the maximum value in the support performance sequence, and use the maximum value as the current design parameter of the adapter; Back propagation is used to train the randomly initialized parameter optimization neural network so that the output value of the parameter optimization neural network is close to the design variable that supports the maximum performance, thereby obtaining the a priori parameter initialization parameters; The parameter optimization neural network is iteratively optimized using the trained uncertainty quantification model and the priori parameter initialization parameters to obtain the optimal model parameters; Determining optimal design parameters of the adapter according to the optimal model parameter parameters, wherein the output of the uncertainty quantification model is the optimal design parameters of the adapter; The iterative optimization process specifically includes: Step 1: Determine the current design parameters of the adapter based on the a priori parameter initialization parameters; Step 2: Perform geometric deformation and mesh generation according to the current design parameters of the adapter to obtain sampling points; Step 3: Use the trained uncertainty quantification model to evaluate the support performance according to the current design parameters of the sampling point and the adapter, and obtain the loss function and objective function of the current sampling point; Step 4: Compare the offspring objective function with the optimal objective function and update it. Use the AdamW optimizer to update the parameters of the parameter optimization neural network to obtain a new generation of parameters. Step 5: Repeat steps 1 to 4 until the iteration stop condition is reached, and save the optimized weight parameters of the parameter optimization neural network after iterative optimization; Use the trained uncertainty quantification model to perform support performance evaluation based on the sampling points and the current design parameters of the adapter, including: Input the current design parameters of the sampling points and the adapter into the trained uncertainty quantification model and calculate the confidence interval of the sampling points; Determine the objective function based on the confidence interval and determine the minimum value of the objective function; The confidence interval is calculated as: ； in, represents the sampling point, represents the uncertainty quantization model after training, represents the number of confidence intervals, represents the design parameters of the adapter at the kth iteration, Represents the current design parameters of the adapter; The calculation formula of the objective function is: ； in, Indicates support performance, Represents the confidence interval The upper bound of Indicates the number of sampling points at the kth iteration.

2. The adapter optimization design method based on parameter optimization neural network according to claim 1, characterized in that: The pre-training of the parameter optimization neural network specifically includes: The AdamW optimizer is used to optimize the parameters according to the loss function of the parameter optimization neural network and the weight parameters of the current parameter optimization neural model to obtain the prior parameter initialization parameters.

3. The adapter optimization design method based on parameter optimization neural network according to claim 1, characterized in that: The training set includes design variables, loads and strain field data of adapter configuration sample points.

4. The adapter optimization design method based on parameter optimization neural network according to claim 1, characterized in that: The adapter is composed of a honeycomb unit cell structure, and the number of design variables of the adapter is 6.

Images