Transformer-enabled intelligent design of high- Q quasi-BIC metasurfaces for molecular vibrational fingerprinting

High-quality-factor (high- Q ) optical resonators are critical for high-sensitivity detection of molecular vibrational fingerprints. Metasurfaces achieve high- Q resonances through quasi-bound states in the continuum (quasi-BICs), making them a promising platform. Deep learning has indicated great potential for designing metasurfaces. However, several limitations remain in the context of high- Q metasurfaces. To overcome these limitations, two core algorithms based on the Transformer architecture are proposed: a Spectrum Vision Transformer (SpecViT) for spectral prediction and a Metasurfaces Vision Transformer (MetaViT) for inverse structural design. Both models are built upon a modified Vision Transformer and significantly improve design efficiency and accuracy when integrated with a spectrum-matching-shifting algorithm. Interpretability of the models is improved through attention mechanism analysis, alleviating the “black-box” nature typical of deep learning models. Using these models, the designed metasurfaces achieved a high Q -factor of approximately 200. By integrating multi-angle multiplexing, these metasurfaces demonstrated high-sensitivity identification of molecular vibrational fingerprints from PMMA and H 2 O. The proposed approach offers a robust intelligent design strategy for high- Q metasurface sensors, promising applications in biomedical detection, trace analysis, and environmental monitoring.