基于物理先验增强代理模型的Vivaldi天线带宽优化设计OA

To address the high computational cost of full-wave electromagnetic simulation,the low proportion of feasible structures,and the limited generalization capability of purely data-driven surrogate models under small-sample conditions in high-dimensional Vivaldi antenna design,a physics-prior-enhanced surrogate modeling and hybrid search method is proposed for bandwidth optimization.An edge-loaded slotted Vivaldi antenna is taken as the study object,and a continuous-discrete hybrid parameterization model is constructed.A Python-High Frequency Structure Simulator(HFSS)automated co-simulation framework is established to realize a closed-loop mapping among structural parameters,geometric models,and electromagnetic performance.In surrogate modeling,physics-prior information related to radiation scale and resonance trend is introduced,and a feature-enhanced Multilayer Perceptron(MLP)is employed for bandwidth prediction.Gaussian Process(GP)residual correction is further incorporated to improve local prediction accuracy.Meanwhile,Monte Carlo dropout(MC dropout)is used to estimate predictive uncertainty,providing confidence information for subsequent search.In the optimization strategy,geometric constraint penalties,dynamic crossover,uncertainty-aware mutation,Bayesian-optimization-based candidate recommendation,and local simulated annealing polishing are integrated to form a multi-strategy collaborative hybrid optimization process.Experimental results show that the proposed method obtains an optimized design with an operating bandwidth of 6.10 GHz under a limited simulation budget.After introducing physics priors and residual correction,the mean absolute error of the surrogate model decreases to 0.073 1,with a determination coefficient of 0.980 0.Compared with the traditional genetic algorithm,the proposed method shows better performance in final bandwidth and high-performance structure search capability.The results indicate that the combination of physics-prior guidance,residual correction,and uncertainty-aware search can improve both modeling accuracy and search efficiency in high-dimensional antenna optimization.