基于迁移学习的超表面设计

Metasurface design based on transfer learning

超表面因具有独特的操控电磁波的功能被广泛使用。但超表面设计需要优化单元结构以产生所需相位分布,这需要大量的时间和计算资源。提出了一种基于迁移学习的超表面设计方法,该方法可以预测复杂超表面单元的相位。使用ResNet34作为预训练模型的迁移学习网络,其中仅使用20000个样本进行训练网络,其准确率达到90%以上。模型实现了预测210×10不同结构的超表面单元的相位,与深度学习网络相比,减少训练样本且提升了准确率。为了验证该方法的有效性和可行性,在实际设计需求下成功设计了一个异常反射在45°角度的相位梯度超表面,并进行了仿真和实验验证。为超表面设计提供了一种新颖高效的方法,并为迁移学习在电磁领域内其他问题上的应用提供了参考。

Metasurface is widely used for its unique ability to manipulate electromagnetic waves.However,metasurface design requires optimization of the cell structure to generate the desired phase distribution,which requires a lot of time and computational resources.Therefore,a metasurface design method based on transfer learning is proposed,which can predict the phase of complex metasurface elements.ResNet34 was used as the transfer learning network of the pre‑training model,in which only 20000 samples were used for the training network,and the accuracy rate reached more than 90%.The model can predict the phase of 210×10 metasurface units with different structures,reducing training samples and improving the accuracy compared with deep learning networks.In order to verify the effectiveness and feasibility of this method,a phase gradient metasurface with abnormal reflection at 45°angle is successfully designed under the actual design requirements,and the simulation and experimental verification are carried out.It provides a novel and efficient method for metasurface design,and provides a reference for the application of transfer learning to other problems in the electromagnetic field.