Redundancy-Guided Active Data Construction for Efficient ML-Assisted Microwave Design

Data-driven surrogate modeling can greatly accelerate microwave component design, but the high cost of full-wave simulations demands efficient data acquisition. Existing sampling strategies often generate redundant samples, limiting model generalization under fixed budgets. This work proposes an active data construction framework driven by a Redundancy Recognition Score (RRS) that quantifies the marginal learning value of each candidate through structural, feature-space, and functional diversity. A closed-loop procedure is established where a Random Forest approximates the RRS landscape and a Differential Evolution optimizer selects high-value samples without requiring prior simulations. Applied to Frequency Selective Surface (FSS) design, the proposed method reduces surrogate generalization error by 43.4\% compared with Latin Hypercube Sampling under the same simulation cost. The framework improves dataset information density and provides a scalable pathway toward efficient ML-assisted microwave design.