Transfer Learning Empowered Multiple-Indicator Optimization Design for Terahertz Quasi-Bound State in the Continuum Biosensors.

Terahertz metasurface biosensors based on the quasi-bound state in the continuum (QBIC) offer label-free, rapid, and ultrasensitive biomedical detection. Recent advances in deep learning facilitate efficient, fast, and customized design of such metasurfaces. However, prior approaches primarily establish one-to-one mappings between structure and optical response, neglecting the trade-offs among key performance indicators. This study proposes a pioneering method leveraging transfer learning to optimize multiple indicators in metasurface biosensor design. For the first time, multiple-indicator comprehensive optimization of the quality (Q) factor, figure of merit (FoM), and effective sensing area (ESA) is achieved. The two-stage transfer learning method pre-trains on low-dimensional datasets to extract shared features, followed by fine-tuning on complex, high-dimensional tasks. By adopting frequency shift as a unified criterion, the contribution ratios of these indicators are quantified as 26.09% for the Q factor, 48.42% for FoM, and 25.49% for ESA. Compared to conventional deep-learning approaches, the proposed method reduces data requirements by 50%. The biosensor designed using this method detects the biomarker homocysteine, achieving detection at the ng µL level, with experimental results closely matching theoretical predictions. This work establishes a novel paradigm for metasurface biosensor design, paving the way for transformative advances in trace biological detection.