A conservative multi-source migration electromagnetic structure optimization design method and related equipment

Abstract

The embodiment of the application provides a conservative multi-source transfer electromagnetic structure optimization design method and related equipment, and belongs to the cross technical field of electromagnetic engineering auxiliary design and artificial intelligence. The method comprises the following steps: establishing a historical source domain database, removing inferior source domains in conflict with a target task through a double-layer screening mechanism based on distribution difference measurement and sorting consistency measurement; fine-tuning the source domain proxy model selected by using initial data of the target domain, and constructing a dynamic weighted integrated proxy model based on uncertainty measurement; using an inverse variance weighting mechanism to explicitly decompose and quantify the intra-model variance and inter-model cognitive divergence; designing a conservative confidence lower bound based on cognitive uncertainty penalty as an acquisition function, and restraining the optimizer by suppressing the predicted value of the high model divergence area; and using a proxy model assisted evolutionary algorithm for iterative optimization and online model updating. The application can effectively utilize historical heterogeneous data to accelerate convergence and avoid negative transfer risk.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of electromagnetic engineering-aided design and artificial intelligence, and in particular to a conservative multi-source migration electromagnetic structure optimization design method and related equipment. Background Technology

[0002] With the rapid development of 6G mobile communication and radar detection technologies, modern radio frequency (RF) front-end systems are placing increasingly stringent performance requirements on electromagnetic devices such as microstrip antennas, cavity filters, and frequency selective surfaces. The design of these electromagnetic structures typically involves dozens or even hundreds of geometric parameters, and their performance indicators and structural parameters exhibit highly nonlinear, non-convex, and multimodal complex mapping relationships. Currently, the mainstream electromagnetic structure design process relies primarily on commercial full-wave electromagnetic simulation software (such as CST Studio Suite and ANSYS HFSS) for verification. However, for complex structures with large electrical dimensions or fine meshing, the time cost of a single high-fidelity simulation is extremely high (ranging from tens of minutes to several hours). Traditional global optimization algorithms typically require thousands or even tens of thousands of iterative evaluations to converge, resulting in excessively long overall design cycles that fail to meet the market demands for rapid product iteration.

[0003] To address this computational bottleneck, Surrogate Model-Based Optimization Aid (SAEA) methods have emerged. These methods utilize machine learning models such as neural networks (NNs) to fit the electromagnetic response surface, replacing expensive physical simulations. However, when dealing with high-dimensional parameter spaces, constructing a high-precision surrogate model requires a large number of training samples, and the data hunger problem remains severe during the cold start phase. Meanwhile, established electromagnetic research institutions typically accumulate vast amounts of historical design data, including optimization trajectories for different frequency bands, substrate materials, and topologies. If this historical knowledge can be reused through transfer learning, the optimization process for new tasks will be greatly accelerated.

[0004] However, existing electromagnetic migration optimization techniques have significant shortcomings in practical applications: 1) Lack of effective source domain screening mechanisms. Existing methods mostly employ single-source migration or blind integration strategies. When there are significant distributional differences (Domain Shift) between the historical task (source domain) and the current design task (target domain) in terms of geometric parameter definitions, frequency band response characteristics, or physical mechanisms, forcibly migrating irrelevant source domain knowledge can mislead the optimization direction, leading to negative transfer, and its convergence speed is even slower than searching from scratch.

[0005] 2) Multi-source model fusion strategies lack quantitative processing of cognitive conflicts. Most existing model fusion techniques rely on simple static weights or global error averaging, making it difficult to distinguish between data noise and cognitive discrepancies between models. In conflicting regions where experts from different source domains give diametrically opposed predictions for the same design area, existing methods often fail to identify the inherent high risks, leading to incorrect estimations of the surrogate model's prediction confidence in these regions and hindering accurate guidance for subsequent optimization searches.

[0006] 3) The optimization strategy suffers from blind optimism. Existing techniques, when using surrogate models to guide optimization, generally tend towards greedy search or blind exploration based on predicted values. In the early stages when data is scarce in the target domain, surrogate models often exhibit significant uncertainty or predictive illusions in unexplored regions. Existing methods lack explicit penalty mechanisms for this cognitive risk and do not impose reasonable geometric constraints on the search radius. This leads to the optimizer easily over-relying on unreliable model predictions, wasting substantial and expensive full-wave simulation resources in ineffective or high-risk parameter spaces, making it difficult to converge to a feasible solution that satisfies engineering constraints within a limited computational budget. Summary of the Invention

[0007] The purpose of this application is to overcome the problems of large sample requirements, low utilization of heterogeneous historical data, and easy occurrence of negative migration in the optimization of complex high-dimensional electromagnetic structures in the prior art. It provides a conservative multi-source migration electromagnetic structure optimization design method, electronic device, storage medium and program product based on uncertainty weighting. Through two-layer source domain screening, dynamic uncertainty weighting integration and conservative acquisition function, it realizes efficient and reliable knowledge reuse and global optimization.

[0008] To achieve the above objectives, one aspect of this application proposes a conservative multi-source migration electromagnetic structure optimization design method, the method comprising: S1: Construct a multi-source historical database and target task modeling. The multi-source historical database contains multiple source domain datasets and corresponding pre-trained proxy models. The target task modeling includes parameterizing the current electromagnetic structure to be designed and obtaining the initial observation dataset. S2: Perform a two-layer source domain screening based on distribution and ranking to select a set of effective source domain expert models compatible with the target task from the multi-source historical database; wherein, the first layer of screening is based on the distribution difference measure of the design space of the source domain and the target domain, and the second layer of screening is based on the ranking consistency measure of the source domain proxy model on the initial data of the target task. S3: Construct a dynamic ensemble model based on uncertainty weighting, fine-tune each model in the effective source domain expert model set using the initial observation dataset, and dynamically calculate the fusion weight based on the local prediction uncertainty of each fine-tuned model at the sample point to be predicted, to obtain the ensemble prediction mean and the total uncertainty after decomposition. S4: Construct a conservative confidence lower bound acquisition function, which is based on the difference between the integrated forecast mean and the total uncertainty weighted by the dynamic risk adjustment factor, and is used to penalize high uncertainty regions during the optimization process; S5: Perform global optimization and online update with the assistance of the surrogate model, solve for the maximum value of the conservative confidence lower bound acquisition function to obtain candidate solutions, verify the real performance of the candidate solutions through simulation, add new samples to the target domain dataset, and update the dynamic ensemble model.

[0009] In some embodiments, the two-layer source domain filtering in step S2 includes: Calculate the maximum mean difference between each source domain dataset and the initial observation dataset of the target domain. If the maximum mean difference exceeds a preset first dynamic threshold, the corresponding source domain is removed. For the source domains that pass the first layer of screening, the corresponding pre-trained proxy model is used to predict the initial observation data of the target domain, and the weighted Kendall rank correlation coefficient between the predicted value and the actual simulation value is calculated. If the weighted Kendall rank correlation coefficient is lower than the preset second dynamic threshold, the corresponding source domain is removed. The first dynamic threshold and / or the second dynamic threshold decrease exponentially with each optimization iteration to achieve a smooth transition from extensive exploration in the early stages of optimization to precise utilization in the later stages.

[0010] In some embodiments, the construction of the uncertainty-weighted dynamic ensemble model in step S3 includes: The fusion weights are calculated using an inverse variance weighting strategy for the sample points to be predicted. , No. The weights of the expert model after fine-tuning Its predicted variance It is directly proportional to the reciprocal; The total uncertainty The explicit decomposition is the sum of within-model variance and between-model discrepancy, where the within-model variance is the weighted average of the predicted variances of each fine-tuned expert model, and the between-model discrepancy is the weighted sum of squares of the differences between the predicted means of each expert model and the ensemble predicted mean.

[0011] In some embodiments, the conservative confidence lower bound acquisition function in step S4 is specifically:

[0012]

[0013] in, This represents the best performance estimate obtained by fusing multi-source knowledge. This is a penalty item for uncertainty; For the number of iterations A function of dynamic decay. The maximum penalty coefficient, This is the attenuation control factor.

[0014] In some embodiments, the construction of the multi-source historical database in step S1, when the design variable dimensions of the source domain and the target domain are inconsistent, includes: Construct an autoencoder neural network containing an encoder and a decoder to map source domain data and target domain data of different dimensions to a common latent space of the same dimension, so as to perform knowledge transfer in a unified space.

[0015] In some embodiments, step S3 further includes a data augmentation strategy based on expert consensus before fine-tuning: The unlabeled candidate point set is input into the effective source domain expert model set, and the consensus degree of all experts on the same point prediction value, i.e. prediction variance, is calculated. Sample points with consensus degree below a preset threshold are selected, and their prediction mean is used as pseudo-labels to construct a virtual dataset and mix it with the real initial observation dataset for pre-training of the source domain expert model.

[0016] In some embodiments, fine-tuning each model in the effective source domain expert model set in step S3 specifically includes: The main network structure parameters of the pre-trained agent model are frozen, and the output layer of the model is iteratively updated using only the initial observation dataset of the target domain. During the fine-tuning process, L2 regularization constraints are introduced to suppress the deviation of the model parameters from the prior knowledge of the source domain.

[0017] In some embodiments, the online update in step S5 includes: After adding new samples to the target domain dataset, each model in the effective source domain expert model set is fine-tuned, and the hyperparameters of inverse variance weighting in subsequent iterations are dynamically adjusted based on the prediction error of each model on the new samples, so as to achieve online adaptive calibration of model weights.

[0018] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0019] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0020] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0021] Compared with the prior art, this application has the following beneficial effects: 1) Explicit Decomposition of Uncertainty: To address the problem that traditional single uncertainty cannot distinguish between data noise and cognitive discrepancies, the total uncertainty is decomposed into intra-model variance and inter-model discrepancy in ensemble modeling, thereby achieving accurate identification and punishment of cognitive discrepancy regions.

[0022] 2) Local confidence dynamic weighting: To address the problem that simple average ensemble cannot reflect the local advantages of experts, a dynamic weighting mechanism based on the inverse of the local prediction variance is proposed, which enables the expert model with the most confidence in the current region to dominate the prediction and improve the local accuracy of the ensemble agent.

[0023] 3) Consensus-driven conservative sampling: In the sampling decision, a conservative strategy of maximizing the mean and reducing the uncertainty penalty is introduced. The region where the multi-source model reaches a consensus is searched first, so as to avoid the problem of the proxy model being overconfident and causing the optimization to fall into a false optimum.

[0024] 4) Adaptive adjustment of risk lower bound parameter: In the construction of confidence lower bound, a risk adjustment term that decays exponentially with the number of iterations is introduced to achieve a smooth switch from strong risk avoidance to fine greedy search.

[0025] 5) Real-time update of inverse variance weights: A mechanism is proposed to update the inverse variance weighting parameters in reverse using the real simulation values ​​after each iteration, so as to realize online adaptive calibration of model weights.

[0026] 6) Rapid initial screening based on distribution differences: The maximum mean difference is introduced as an initial screening indicator to quickly eliminate source domains with excessive distribution offset and compress redundant knowledge space.

[0027] 7) Precise screening based on ranking ability: A ranking ability evaluation index based on weighted Kendall coefficient is proposed to accurately measure the guiding value of the model for optimization direction.

[0028] 8) Adaptive two-layer dynamic screening: Construct a two-layer screening architecture and introduce a dynamic threshold that decays exponentially with each iteration round to achieve an adaptive transition from broadly absorbing prior knowledge in the early stage to strictly retaining experts in the later stage.

[0029] 9) Unifying the Dimensions of Heterogeneous Data: A mapping mechanism based on an autoencoder is proposed to embed heterogeneous data into a unified low-dimensional latent space, solving the problem of inconsistent design variable dimensions between the source domain and the target domain.

[0030] 10) Consensus-based data augmentation strategy: A method for screening unlabeled samples and generating virtual data based on the consensus of the source domain expert set is proposed. This method expands the training set without introducing full-wave simulation, significantly alleviating the bottleneck of data scarcity.

[0031] 11) Regularization of source domain knowledge preservation: Introduce a parameter deviation penalty term when updating the model to constrain the model from deviating from the source domain prior while fitting new data.

[0032] 12) Lightweight output layer fine-tuning: A matching strategy is proposed to freeze the underlying feature extraction network and only update the output mapping layer to achieve fast domain adaptation with extremely low training cost.

[0033] 13) Supports incremental fine-tuning updates: To address the problem that offline trained models cannot adapt to new sample features, new samples are added to the training set in real time after full-wave simulation and the output layer is fine-tuned, so that the accuracy of the surrogate model continues to approach that of the real simulator as the optimization process progresses.

[0034] 14) Provide a modular and general framework: This method adopts a modular design, supports user-defined distribution metrics, ranking metrics, integration and fusion strategies and optimization solvers, has good versatility and scalability, and can be adapted to different physical scenarios and data conditions. Attached Figure Description

[0035] Figure 1 This is a flowchart of the conservative multi-source migration electromagnetic structure optimization design method provided in the embodiments of this application; Figure 2 This is a framework diagram of the conservative multi-source migration electromagnetic structure optimization design method based on uncertainty weighting provided in the embodiments of this application; Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0038] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0039] 1) Conservative Confidence Lower Bound: The conservative confidence lower bound is a function used to guide the global heuristic optimization algorithm in selecting the next generation of evaluation samples during surrogate model-assisted multi-source transfer optimization. Mathematically, this function is composed of the predicted mean of the multi-source dynamic ensemble model and a subtractive penalty term based on the total cognitive uncertainty. By introducing a subtractive penalty, the conservative confidence lower bound applies numerical suppression to high-risk regions where there are conflicts in multi-source expert opinions or where the model prediction variance is large. This constrains the optimization algorithm to prioritize accurate searches within the high-confidence safe region where the multi-source models reach consensus, effectively avoiding the risk of local oscillations and negative transfer caused by surrogate model prediction bias, and ensuring the reliability of the optimization process for complex high-dimensional electromagnetic structures.

[0040] To address the problems of high cost of high-dimensional simulation, susceptibility to negative transfer learning in traditional electromagnetic structure design, and overconfidence during optimization, this application provides a conservative multi-source transfer electromagnetic structure optimization design method, electronic device, storage medium, and program product based on uncertainty weighting. The solution includes: First, establishing a historical source domain database and eliminating inferior source domains that conflict with the target task through a two-layer screening mechanism based on distribution difference and ranking consistency measures; next, fine-tuning the selected source domain surrogate models using initial data from the target domain and constructing a dynamic weighted ensemble surrogate model based on uncertainty measures, explicitly decomposing and quantifying the intra-model variance and inter-model cognitive divergence using an inverse variance weighting mechanism; subsequently, designing a conservative confidence lower bound based on cognitive uncertainty penalty as the acquisition function, constraining the optimizer by suppressing predicted values ​​in high model divergence regions; finally, using a surrogate model-assisted evolutionary algorithm for iterative optimization and online model updating. This application can effectively utilize historical heterogeneous data to accelerate convergence, avoid the risk of negative transfer, and significantly improve the efficiency and reliability of complex electromagnetic structure design.

[0041] The conservative multi-source migration electromagnetic structure optimization design method provided in this application relates to the interdisciplinary field of electromagnetic engineering-aided design and artificial intelligence. This method can be applied to terminals, servers, or software running on either. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the conservative multi-source migration electromagnetic structure optimization design method, but is not limited to the above forms.

[0042] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0043] like Figure 1 As shown, this embodiment provides a conservative multi-source migration electromagnetic structure optimization design method, which specifically includes the following steps: Step S101: Construct a multi-source historical database and target task modeling. The multi-source historical database contains multiple source domain datasets and corresponding pre-trained surrogate models. Target task modeling includes parameterizing the current electromagnetic structure to be designed and obtaining the initial observation dataset. Step S102: Perform a two-layer source domain screening based on distribution and ranking to select a set of effective source domain expert models compatible with the target task from a multi-source historical database; wherein, the first layer of screening is based on the distribution difference measure of the design space of the source domain and the target domain, and the second layer of screening is based on the ranking consistency measure of the source domain proxy model on the initial data of the target task. Step S103: Construct a dynamic ensemble model based on uncertainty weighting. Fine-tune each model in the effective source domain expert model set using the initial observation dataset. Based on the local prediction uncertainty of each fine-tuned model at the sample point to be predicted, dynamically calculate the fusion weight to obtain the ensemble prediction mean and the total uncertainty after decomposition. Step S104: Construct a conservative confidence lower bound acquisition function. The acquisition function is constructed based on the difference between the ensemble forecast mean and the total uncertainty weighted by the dynamic risk adjustment factor, and is used to penalize the high uncertainty region during the optimization process. Step S105: Perform global optimization and online update with surrogate model assistance, solve for the maximum value of the conservative confidence lower bound acquisition function to obtain candidate solutions, verify the real performance of candidate solutions through simulation, add new samples to the target domain dataset, and update the dynamic ensemble model.

[0044] The solutions of the embodiments of this application will be described in detail below with reference to the accompanying drawings and specific application examples.

[0045] To overcome the problems of large sample requirements, low utilization of heterogeneous historical data, and susceptibility to negative migration in existing technologies for optimizing complex high-dimensional electromagnetic structures, this embodiment provides a conservative multi-source migration electromagnetic structure optimization design method based on uncertainty weighting. This method achieves efficient and reliable knowledge reuse and global optimization through two-layer source domain screening, dynamic uncertainty weighted integration, and a conservative acquisition function. See also... Figure 2 The method specifically includes the following five steps: Step S1: Construct a multi-source historical database and model the target task.

[0046] S1-1: Parametric Modeling of Electromagnetic Structures In this step, the electromagnetic structure to be designed (such as a broadband microstrip patch antenna or a multilayer frequency selective surface) is first parametrically modeled. The design variables are then defined as vectors. ,in For variable dimensions. Variable types can include continuous variables (physical dimensions, dielectric constant, etc.) and discrete variables (number of layers, switch states, etc.).

[0047] Define the optimization objective function For example, maximizing the gain of a specific frequency band, minimizing the S11 parameter (return loss), or optimizing the axial ratio bandwidth. If it is a multi-objective problem, it can be transformed into a weighted single objective or processed using a Chebyshev aggregation function.

[0048] The target task is denoted as Its initial observation dataset is The sample size was obtained through Latin hypercube sampling (LHS) and calculated using full-wave simulation software. Typically small, for example, 10-50.

[0049] S1-2: Source Domain Database Construction Collect and organize data from past electromagnetic design projects to construct a source domain set. ,in Number of source tasks. Number of source domains. Includes historical datasets and the agent model pre-trained based on this data .

[0050] To address situations where the dimensions of design variables in the source domain are inconsistent with those in the target domain, this embodiment proposes a common latent space projection technique based on an autoencoder. The specific implementation steps are as follows: Construct a common latent space projection technique containing an encoder... and decoder The neural network structure. For a dimension of Source domain data and dimension are Target domain data Each input layer is constructed separately and mapped to a common latent space of the same dimension. .

[0051] S1-3: Construction of the Source Domain Proxy Model Using the source domain dataset prepared in step S1-2 Train the agent model for each source task, and denote the first... The proxy model corresponding to each source task is: The proxy model It must possess the ability to quantify uncertainty, that is, for any given input design variables... The model can not only output the predicted mean In order to characterize the estimation of electromagnetic properties, it must also be able to output the prediction variance. This is used to characterize the confidence level of the prediction result. In a specific embodiment of the present invention, the surrogate model is constructed using Gaussian Process Regression (GPR) or Bayesian Neural Network (BNN).

[0052] Step S2: Perform a two-layer source domain screening based on distribution and sorting.

[0053] To reduce the risk of negative migration, this embodiment designs a two-layer source domain screening mechanism.

[0054] S2-1: First-level screening – Measurement of distribution differences Step S2-1 aims to assess the degree of overlap between the source domain design space and the target domain design space.

[0055] Compute source domain dataset Input feature distribution With the target domain dataset Input feature distribution The distribution difference metric is used to measure the difference between data points. Specific implementation techniques for this distribution difference metric include, but are not limited to: Maximum Mean Discrepancy (MMD), Körbeck-Leibler Divergence (KL Divergence), Jensen-Shannon Divergence (JS Divergence), Wasserstein Distance, or domain discriminator scores based on adversarial networks. This invention selects Maximum Mean Discrepancy as the specific implementation. MMD utilizes kernel function techniques to map data to a Reproducing Kernel Hilbert Space (RKHS), and its calculation formula is as follows:

[0056] in, and These represent the number of samples in the source domain dataset and the target domain dataset, respectively. This represents a feature mapping function used to map the original input space to the reproducing kernel Hilbert space. Using Gaussian kernel function If the calculated index value exceeds the preset distribution threshold, it is determined that the feature distribution of the source domain is inconsistent with that of the target domain, and it is removed from the candidate set.

[0057] S2-2: Second-level screening – Ranking consistency measurement Step S2-2 aims to evaluate whether the source domain model has mastered the judgment logic for the target task. This is achieved using the model trained in the source domain. Predict existing samples in the target domain and calculate a ranking consistency metric between the predicted value vector and the actual simulated value vector.

[0058] The specific implementation techniques of the ranking consistency metric include, but are not limited to: Weighted Kendall's rank correlation coefficient. Spearman's rank correlation coefficient Top-K accuracy (Precision@K) or Normalized Discount Cumulative Gain (NDCG).

[0059] This embodiment uses the weighted Kendall's rank correlation coefficient. This is the specific implementation. Its calculation formula is as follows:

[0060] in, The index for the sample pairs in the target domain. ; Indicates sample and samples The actual performance values ​​obtained through electromagnetic simulation; Represents the source domain proxy model For the sample and samples The predicted value; This is a sign function; it outputs 1 when the input is greater than 0, -1 when the input is less than 0, and 0 otherwise. To assign sample pairs The weighting coefficients are determined based on the actual performance ranking of the samples. Samples with better performance receive larger weighting coefficients to strengthen the assessment of the accuracy of the ranking of top samples. If the calculated index value is lower than a preset threshold, the source domain model is deemed unable to accurately reflect the superiority or inferiority relationship of the target task and is removed from the candidate set.

[0061] To adapt to changes in data volume during the optimization process, this embodiment designs a dynamic threshold filtering mechanism. This embodiment defines the threshold based on the iteration rounds. Changing dynamic threshold :

[0062] In the early stages of optimization ( Smaller), threshold It is wider, allowing for the introduction of more potentially valuable source domains to alleviate the cold start; as iterations proceed ( The threshold is increased, and then tightened exponentially, retaining only high-confidence experts who are highly consistent with the target task, thus achieving a gradual transition from 'broad exploration' to 'precise utilization'. Finally, the set of source models retained after the two-layer screening is denoted as . .

[0063] Step S3: Construct a dynamic ensemble model based on uncertainty weighting.

[0064] S3-1: Target Domain Data Augmentation To alleviate the problem of extremely scarce real samples in the initial stage of the target domain before building the ensemble model, this embodiment introduces a data augmentation strategy based on expert consensus.

[0065] Latin hypercube sampling is used to generate a large set of unlabeled candidate points within the design space. Input this point set into the filtered source domain expert set. In this process, the consensus among all experts on the same predicted value is calculated, i.e., the prediction variance. :

[0066] Set consensus threshold Select the one that satisfies The sample points represent low-risk areas agreed upon by multiple experts. Their predicted mean is used as pseudo-labels to construct the dataset. and with real datasets The mixture is used for pre-training of subsequent surrogate models, thereby expanding the training samples without increasing the cost of full-wave simulation.

[0067] S3-2: Domain Adaptation Fine-tuning of the Source Domain Expert Model Despite the filtering in step S2, the remaining source domain models It is still trained based on historical data distribution, and directly applying it to the target domain results in systematic bias. In this step, the initial small sample dataset of the target domain obtained in step S1 is used. For sets Each source domain expert model is fine-tuned, and L2 regularization constraints are introduced during the fine-tuning process. First, the main network structure or kernel parameters of the source model are kept unchanged, and only the following parameters are used: The output layer weights of the model are iteratively updated a few times to make its predicted distribution approximate the true response surface of the target domain. The fine-tuned expert model is denoted as... It is no longer a static historical prior, but an adaptive expert that adapts to the characteristics of the current task.

[0068] S3-3: Inverse Variance Weighted (IVW) Mechanism For any sample point to be predicted in the design space Calculate each expert model in the effective expert set. Fusion weights .

[0069] This invention employs an inverse variance weighting strategy to calculate the fusion weights. :

[0070] in, To prevent numerically stable terms with a denominator of zero (e.g., taking...) ).

[0071] This mechanism establishes an adaptive fusion principle based on local confidence: within a specific local feature space, expert models with smaller prediction variance and higher confidence will be assigned larger weight coefficients, thus dominating the integrated prediction results for that region; conversely, the weights of models with larger prediction variance will be automatically suppressed. This mechanism ensures that the integrated model can dynamically and optimally utilize multi-source knowledge, minimizing the interference of uncertainty from a single model.

[0072] S3-4: Decomposition of Total Uncertainty The mean of the integrated predictions is: This invention integrates the total uncertainty of the model. It is decomposed into two parts: within-model variance and between-model discrepancy.

[0073]

[0074] Intra-model variance reflects the average level of hesitation among all experts. Inter-model disagreement reflects the degree of contention among experts. This term will increase significantly if expert A predicts a high value while expert B predicts a low value. This usually indicates severe epistemic uncertainty in this area, making it a potentially high-risk region.

[0075] Step S4: Construct a conservative confidence lower bound acquisition function.

[0076] In multi-source migration scenarios, unknown regions often imply cognitive risks caused by differences in the distribution of source domain data. Therefore, this embodiment adopts an acquisition function based on the principle of conservatism, namely the Conservative Lower Bound (CLB). This function no longer pursues the maximization of the potential upper limit of returns, but seeks to maximize the performance lower limit in the worst case.

[0077] S4-1: Definition of Conservative Confidence Lower Bound Construct a get function of the following form The optimization objective is to find the maximum value of the function:

[0078]

[0079] in: This represents the best performance estimate obtained by fusing multi-source knowledge. This is an uncertainty penalty term. Unlike the plus sign in the Upper Confidence Bound (UCB) strategy, a minus sign is used here. This means that if a region has a high predicted value but the model has high uncertainty about it, the optimization algorithm will lower its evaluation of that point. This guides the search towards high-confidence regions where multi-source expert models have reached consensus and have excellent prediction performance, effectively avoiding the convergence oscillation risk caused by prediction biases in surrogate models. Furthermore, to balance exploration and development... Set to vary with the number of iterations The dynamically decaying function makes the uncertainty decrease in the later stages of optimization as more real data from the target domain accumulates. As the variance gradually decreases, the algorithm performs a more refined search in regions with larger variances, eventually converging to the globally theoretical optimal solution. The maximum penalty coefficient is used to define the maximum penalty applied to cognitive uncertainty during the initial optimization phase of the algorithm. This is a decay control factor used to control the rate at which the conservative penalty decays with each iteration.

[0080] Step S5: Perform global optimization and online update with the assistance of the proxy model.

[0081] S5-1: Subproblem Optimization because It is a highly nonlinear, multimodal function, and its analytical gradient is difficult to obtain. This step uses a global heuristic optimization algorithm to solve it and find the best candidate solution for the current iteration step. The specific implementation techniques of the global heuristic optimization algorithm include, but are not limited to: Covariance Matrix Adaptive Evolution Strategy (CMA-ES), Differential Evolution (DE), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Simulated Annealing (SA), or Ant Colony Optimization (ACO). This embodiment considers the advantages of CMA-ES in handling continuous variables and ill-conditioned conditions, and therefore selects CMA-ES as the specific implementation. The algorithm finds the point that maximizes the conservative confidence lower bound:

[0082] S5-2: Simulation Verification Will The input is used in electromagnetic simulation software for precise calculations to obtain the actual response. .

[0083] S5-3: Online Learning New sample Add to target domain dataset The source domain expert models were fine-tuned using the updated target domain dataset, and then all source models were recalculated. The prediction error at the point is used to dynamically adjust the hyperparameters of IVW in subsequent steps. With... Increased data volume, uncertainty terms The value will gradually decrease, and the algorithm will automatically transition from conservatively relying on prior data to relying on real data, eventually converging to the global optimum.

[0084] In summary, existing surrogate model-assisted electromagnetic structure optimization techniques typically employ a single surrogate model to fit the objective function, combined with traditional heuristic strategies for global search. When utilizing historical data, existing techniques often employ direct data merging or simple one-to-one parameter transfer strategies. When dealing with multi-source heterogeneous data with significantly different distributions, this approach, lacking an applicability evaluation and screening mechanism for source domain data, easily introduces inferior prior knowledge that conflicts with the physical laws of the target task, leading to negative transfer and causing optimization convergence speeds that are even slower than random searches. Furthermore, existing techniques generally follow an optimistic exploration principle in the face of uncertainty in multi-model integration and optimization decision-making, ignoring potential cognitive discrepancies between multi-source models. This causes the optimizer to easily consume a large amount of expensive full-wave simulation resources in regions where surrogate model predictions are distorted.

[0085] Compared with the existing technologies, this application proposes a conservative multi-source transfer optimization method based on uncertainty weighting. Its significant differences are: 1) This application rigorously assesses the source tasks from two dimensions: feature distribution space overlap and performance ranking capability, thus preventing the interference of inferior data on the target task from the source. 2) This application can dynamically allocate weights based on the prediction variance of each expert model in a local region and explicitly decompose the total uncertainty into intra-model variance and inter-model disagreement, thereby accurately quantifying the degree of conflict in multi-source expert opinions. 3) This application uses inter-model disagreement as a risk penalty term to construct an acquisition function, constraining the global heuristic optimization algorithm to avoid regions with high cognitive risk and prioritizing searches within credible regions where multi-source models have reached consensus.

[0086] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0087] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0088] Please see Figure 3 , Figure 3 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 302 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 302 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 302 and is called and executed by the processor 301 using the methods described above in the embodiments of this application. Input / output interface 303 is used to implement information input and output; The communication interface 304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 305 transmits information between various components of the device (e.g., processor 301, memory 302, input / output interface 303, and communication interface 304); The processor 301, memory 302, input / output interface 303, and communication interface 304 are connected to each other within the device via bus 305.

[0089] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0090] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0091] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0092] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0093] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0094] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0095] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0096] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0098] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0099] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0101] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0103] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A conservative multi-source migration electromagnetic structure optimization design method, characterized in that, The method includes the following steps: S1: Construct a multi-source historical database and target task modeling. The multi-source historical database contains multiple source domain datasets and corresponding pre-trained proxy models. The target task modeling includes parameterizing the current electromagnetic structure to be designed and obtaining the initial observation dataset. S2: Perform a two-layer source domain screening based on distribution and ranking to select a set of effective source domain expert models compatible with the target task from the multi-source historical database; wherein, the first layer of screening is based on the distribution difference measure of the design space of the source domain and the target domain, and the second layer of screening is based on the ranking consistency measure of the source domain proxy model on the initial data of the target task. S3: Construct a dynamic ensemble model based on uncertainty weighting, fine-tune each model in the effective source domain expert model set using the initial observation dataset, and dynamically calculate the fusion weight based on the local prediction uncertainty of each fine-tuned model at the sample point to be predicted, to obtain the ensemble prediction mean and the total uncertainty after decomposition. S4: Construct a conservative confidence lower bound acquisition function, which is based on the difference between the integrated forecast mean and the total uncertainty weighted by the dynamic risk adjustment factor, and is used to penalize high uncertainty regions during the optimization process; S5: Perform global optimization and online update with the assistance of the surrogate model, solve for the maximum value of the conservative confidence lower bound acquisition function to obtain candidate solutions, verify the real performance of the candidate solutions through simulation, add new samples to the target domain dataset, and update the dynamic ensemble model.

2. The method according to claim 1, characterized in that, The two-layer source domain filtering described in step S2 includes: Calculate the maximum mean difference between each source domain dataset and the initial observation dataset of the target domain. If the maximum mean difference exceeds a preset first dynamic threshold, the corresponding source domain is removed. For the source domains that pass the first layer of screening, the corresponding pre-trained proxy model is used to predict the initial observation data of the target domain, and the weighted Kendall rank correlation coefficient between the predicted value and the actual simulation value is calculated. If the weighted Kendall rank correlation coefficient is lower than the preset second dynamic threshold, the corresponding source domain is removed. The first dynamic threshold and / or the second dynamic threshold decrease exponentially with each optimization iteration to achieve a smooth transition from extensive exploration in the early stages of optimization to precise utilization in the later stages.

3. The method according to claim 1, characterized in that, The construction of the dynamic ensemble model based on uncertainty weighting in step S3 includes: The fusion weights are calculated using an inverse variance weighting strategy for the sample points to be predicted. , No. The weights of the expert model after fine-tuning Its predicted variance It is directly proportional to the reciprocal; The total uncertainty The explicit decomposition is the sum of within-model variance and between-model discrepancy, where the within-model variance is the weighted average of the predicted variances of each fine-tuned expert model, and the between-model discrepancy is the weighted sum of squares of the differences between the predicted means of each expert model and the ensemble predicted mean.

4. The method according to claim 1, characterized in that, The conservative confidence lower bound acquisition function mentioned in step S4 is specifically as follows: in, This represents the best performance estimate obtained by fusing multi-source knowledge. This is a penalty item for uncertainty; For the number of iterations A function of dynamic decay. The maximum penalty coefficient, This is the attenuation control factor.

5. The method according to claim 1, characterized in that, Step S1, which involves constructing a multi-source historical database, includes the following when the design variable dimensions of the source domain and the target domain are inconsistent: Construct an autoencoder neural network containing an encoder and a decoder to map source domain data and target domain data of different dimensions to a common latent space of the same dimension, so as to perform knowledge transfer in a unified space.

6. The method according to claim 1, characterized in that, Step S3, prior to fine-tuning, further includes a data augmentation strategy based on expert consensus: The unlabeled candidate point set is input into the effective source domain expert model set, and the consensus degree of all experts on the same point prediction value, i.e. prediction variance, is calculated. Sample points with consensus degree below a preset threshold are selected, and their prediction mean is used as pseudo-labels to construct a virtual dataset and mix it with the real initial observation dataset for pre-training of the source domain expert model.

7. The method according to claim 1, characterized in that, Step S3 involves fine-tuning each model in the set of effective source domain expert models, specifically including: The main network structure parameters of the pre-trained agent model are frozen, and the output layer of the model is iteratively updated using only the initial observation dataset of the target domain. During the fine-tuning process, L2 regularization constraints are introduced to suppress the deviation of the model parameters from the prior knowledge of the source domain.

8. The method according to claim 1, characterized in that, The online update mentioned in step S5 includes: After adding new samples to the target domain dataset, each model in the effective source domain expert model set is fine-tuned, and the hyperparameters of inverse variance weighting in subsequent iterations are dynamically adjusted based on the prediction error of each model on the new samples, so as to achieve online adaptive calibration of model weights.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images