Deep Learning for Electromagnetic Metamaterial Inverse Design

Designing metamaterials involves computationally intensive tasks, resulting in a time-consuming design process. A deep learning approach is proposed to generate metamaterial designs directly from the input S-parameter diagram. Furthermore, to address the challenge of creating the input diagram, a fine-tuned Large Language Model (LLM) generates the S-parameter diagram based on user input prompts. Compared to traditional neural network methods, the use of an image-based deep learning network enhances the architecture by minimizing the pre and post-processing steps required to transform a metamaterial unitcell image into an numerical array, and vice versa. Simulation results confirm the agreement between the generated metamaterial and the desired input response.