Model Agnostic Data Enhancement Algorithm in 2D Chiral Metamaterials: From Prediction to Inverse Design

In this paper, we introduce a powerful model agnostic data enhancement (MADE) learning algorithm to realize the efficient prediction and rapid inverse design of 2D diffractive chiral metamaterials. In order to vividly demonstrate the performance of our MADE algorithm, multi-shaped chiral structures with various geometrical parameters (e.g., width, separation space, module length, depths) and material components are selected. To generate the training data for MADE model, we utilize the conventional method of rigorous coupled wave analysis (RCWA) to capture the chiroptical response of metamaterials. Moreover, data enhancement and multi-task learning strategy constitute two key sections of MADE algorithm. First, a forward prediction network replying on data enhancement is built to generate large amount of "pseudo data" to reinforce the original training dataset that is much less than previous work, which can also perform fast prediction of the metamaterials’ circular dichroism response. Additionally, multi-task learning strategy is exploited in MADE algorithm to address the issue that various geometric parameters apparently affect the chiroptical response in a different manner, enhancing the model versatility. The comprehensive results demonstrate that our MADE algorithm can not only carry out the forward prediction for metamaterials’ chiroptical response with a significantly-fast computational speed that is 4 orders of magnitude faster than RCWA, but equips with fast inverse design function for on-demand metamaterials, which also can be applied to contrive new structures with desired optical response. This study set out to assess the feasibility of deep learning method in the large-scale, efficient design of 2D diffractive chiral metamaterials and more sophisticated photonic devices on chip.