Objective To address the problem that deep-learning-based optimization of metasurfaces relies heavily on large-scale full-wave simulation data, thereby limiting optimization efficiency, the dataset construction methods, deep learning model structure construction methods, deep learning model training methods, and optimization methods for metasurface structure design scenarios are investigated.

Method First, a training-set sampling method based on sample importance is developed. By evaluating the gradient information of the loss function relative to the samples, this method strategically identifies and selects highly informative data points, significantly reducing the required sample volume while improving the model's accuracy in estimating the electromagnetic responses of metasurface structures. Second, a multimodal deep learning model is constructed to simultaneously extract and integrate features from both vectorized structural parameters and pixelated patterns. Through a systematic feature fusion mechanism, the structural representation capability is enhanced, further improving the response-estimation performance. Third, a novel training method leveraging the out-of-distribution (OOD) generalization property of the deep learning model is proposed. This strategy utilizes the model's intrinsic generalization capabilities to synthesize and introduce low-fidelity response samples outside the initial training distribution, dynamically expanding the feature space and thereby reducing the necessary scale of the high-fidelity training sample set. Finally, instead of relying on a globally accurate model with high computational training costs, an efficient optimization method for metasurface structures is proposed. This approach utilizes a coarse deep learning model trained with a strictly limited sample size, operating in conjunction with an iterative refinement mechanism to guide the optimization process.

Results Numerical results demonstrate the high efficiency of this data-driven framework. Specifically, while the baseline performance of the deep learning models is rigorously maintained, the respective implementations of the proposed dataset construction method, multimodal model architecture, and OOD training strategy each reduce the number of full-wave simulation samples required for initial training by 30%–50%. This substantial reduction directly alleviates the computational burden associated with generating high-fidelity datasets. Furthermore, during the practical optimization phase, it is demonstrated that the proposed algorithm based on the coarse deep learning model achieves rapid convergence. Metasurface structures exhibiting excellent electromagnetic performance can be successfully designed and synthesized, requiring only dozens of additional iterations and full-wave simulation validations, proving the method's capability to bypass the reliance on highly precise, computationally expensive surrogate models.

Conclusion A low-data-dependency system framework is detailed that encompasses the process of “data-model-training-application”. Through targeted restructuring at four different levels, it systematically addresses the challenge of data scale limitations that hinder the efficiency of metasurface design. As a result, it provides general methodological support and a paradigm for the low-cost application of deep learning in the field of electromagnetic engineering.