Deep-learning-assisted intelligent design of terahertz hybrid-functional metasurfaces with freeform patterns

• A generator network with a reconstructed loss function is developed to improve the reliability of predicting MS structures. • An enhanced dataset including hybrid-mode and freeform-pattern MSs is created for designing multifunctional MSs. • A neural network is improved by mode labeling to identify different functions in forward and inverse training process. Metasurfaces (MSs) possess powerful abilities to precisely manipulate electromagnetic (EM) waves. However, traditional metasurface design relies on time-consuming trial-and-error methods, and the design process becomes increasingly complex and challenging, particularly with the growing demand to integrate multiple functions into a single MS . To overcome these challenges, this work develops an innovative deep Unet++ conditional generative adversarial network (GAN), named Deep UCGAN++, for designing and characterizing hybrid-functional MSs with freeform patterns. The developed deep learning (DL) network model achieves not only accurate forward prediction of the EM response for given MSs, but also enables the inverse design of candidate MSs based on the target of user-defined functionalities. By constructing and training a hybrid-mode database of freeform MS, we effectively overcome the challenges of the limited design freedom and the deficiency of model generalization ability in DL-assisted MSs. This work offers a robust, high-accuracy yet time-efficient method for the generation of hybrid-functional MSs, paving the way for the further development of reconfigurable multifunctional MSs.