Optical metasurfaces with pronounced spectral characteristics are promising for sensor applications. Currently, deep learning (DL) offers a rapid manner to design various metasurfaces. However, conventional DL models are usually assumed as black boxes, which is difficult to explain how a DL model learns physical features, and they usually predict optical responses of metasurfaces in a fuzzy way. This makes them incapable of capturing critical spectral features precisely, such as high quality (Q) resonances, and hinders their use in designing metasurface sensors. Here, a transformer-based explainable DL model named Metaformer for the high-intelligence design, which adopts a spectrum-splitting scheme to elevate 99% prediction accuracy through reducing 99% training parameters, is established. Based on the Metaformer, all-dielectric metasurfaces based on quasi-bound states in the continuum (Q-BIC)for high-performance metasensing are designed, and fabrication experiments are guided potently. The explainable learning relies on spectral position encoding and multi-head attention of meta-optics features, which overwhelms traditional black-box models dramatically. The meta-attention mechanism provides deep physics insights on metasurface sensors, and will inspire more powerful DL design applications on other optical devices.
Read full abstract