Abstract: Objectives 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 metasurcae structure design scenarios are investigated. Methods First, a training-set sampling method based on sample importance is developed, by which the number of samples required for training is reduced while the model performance in estimating the electromagnetic responses of metasurface structures is improved. Second, a multimodal deep learning model is constructed to simultaneously extract features form vectorized structural parameters and pixelated patterns of metasurface structures, by which the response-estimation performance is improved. Third, a training method that leverages the out-of-distribution generalization property of the deep learning model is proposed, by which low-fidelity samples are generated using the model’s own generalization capability and the scale of the high-fidelity training sample set is thereby reduced. Finally, instead of using an accurate model with high computational cost, an optimization method for metasurface structure is proposed in which a coarse deep learning model trained with a limited sample size is used. Results Numerical results demonstrate that, while the performance of the deep learning model is maintained, the proposed dataset construction method, multimodal model construction method, and training method can reduce the number of full-wave simulation samples required for training by 30%–50%, thereby substantially reducing the full-wave simulations required for optimizing metasurface structures. In addition, it is demonstrated that, by the proposed optimization method based on the coarse deep learning model, metasurface structures with excellent performance can be designed through dozens of iterations and full-wave simulation validations. Conclusions A low-data-dependency system framework is detailed that encompasses the entire closed-loop 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 superurface design. As a result, it provides a general methodological support and paradigm for the low-cost application of deep learning in the field of electromagnetic engineering.