Cross-domain heterogeneous metasurface inverse design based on a transfer learning method

In this Letter, a transfer learning method is proposed to complete design tasks on heterogeneous metasurface datasets with distinct functionalities. Through fine-tuning the inverse design network and freezing the parameters of hidden layers, we successfully transfer the metasurface inverse design knowledge from the electromagnetic-induced transparency (EIT) domain to the three target domains of EIT (different design), absorption, and phase-controlled metasurface. Remarkably, in comparison to the source domain dataset, a minimum of only 700 target domain samples is required to complete the training process. This work presents a significant solution to lower the data threshold for the inverse design process and provides the possibility of knowledge transfer between different domain metasurface datasets.