Objective To address the issue that deep-learning-based metasurface optimization relies heavily on large-scale full-wave simulation data, thereby limiting optimization efficiency, this study investigates dataset construction methods, deep learning model architecture design methods, deep learning model training strategies, and optimization approaches for metasurface structure design scenarios.

Method  First, a training-set sampling method based on sample importance is developed. By evaluating the gradient information of the loss function with respect to individual samples, this method strategically identifies and selects highly informative data points, significantly reducing the required sample size while improving the accuracy of electromagnetic response prediction for metasurface structures. Second, a multimodal deep learning model is constructed to simultaneously extract and integrate features from both vectorized structural parameters and pixelated pattern representations. Through a systematic feature fusion mechanism, the model enhances structural representation capability and further improves electromagnetic response prediction performance. Third, a novel training strategy that exploits the out-of-distribution (OOD) generalization capability of the deep learning model is proposed. This strategy leverages the intrinsic generalization ability of the model to synthesize and incorporate low-fidelity response samples outside the original training distribution, dynamically expanding the feature space and thereby reducing the required scale of the high-fidelity training dataset. Finally, instead of relying on a globally accurate model that incurs substantial computational training costs, an efficient optimization method for metasurface structure design is proposed. This approach employs a coarse deep learning model trained on a strictly limited dataset and combines it with an iterative refinement mechanism to guide the optimization process.

Results Numerical results demonstrate the high efficiency of the proposed data-driven framework. Specifically, while rigorously maintaining the baseline performance of the deep learning models, the proposed dataset construction method, multimodal model architecture, and OOD-based training strategy each reduce the number of full-wave simulation samples required for initial training by 30%–50%. This substantial reduction significantly alleviates the computational burden associated with generating high-fidelity datasets. Furthermore, during the practical optimization stage, the proposed optimization algorithm based on a coarse deep learning model is shown to achieve rapid convergence. Metasurface structures with excellent electromagnetic performance can be successfully designed and synthesized using only dozens of additional iterative updates and full-wave simulation validations. These results demonstrate the capability of the proposed method to eliminate the dependence on highly accurate yet computationally expensive surrogate models.

Conclusion A low-data-dependency system framework covering the complete process of "data−model−training−application" is established. Through targeted restructuring at four different levels, it systematically addresses the challenge of limited data scale that constrains the efficiency of metasurface design. As a result, it provides a general methodological foundation and practical paradigm for the low-cost application of deep learning techniques in the field of electromagnetic engineering.