High-performance Chiral Metasurface Sensors Optimized by a Target-driven Active Learning Framework

• A hybrid target-driven active learning framework is proposed for the efficient inverse design of chiral metasurfaces, combining uncertainty-aware reliability filtering with global optimization. • The core active learning engine uses an ensemble of CNNs to quantify prediction uncertainty, guiding the selection of structures for costly simulation. • This approach reduces the design cycle from months to 10 days . • The framework successfully designs meta-units with strong circular dichroism (CD > 0.7) and achieves excellent regression performance with a validation mean squared error (MSE) of 2.1 × 10 -4 and an R 2 of 0.9969. In recent years, with the rise of artificial intelligence (AI) algorithms, the development of nanophotonics has been significantly promoted, and shown great potential in material design, device optimization and performance prediction. To address the inefficiency of traditional nanophotonic design methods and the challenges of multi-dimensional parameter optimization, we propose a target-driven active learning framework that synergistically integrates uncertainty-aware reliability filtering with efficient global optimization for the inverse design of chiral metasurface sensors. The framework employs an ensemble of Convolutional Neural Networks (CNNs) as a surrogate model and leverages their prediction disagreement as an uncertainty metric to guide the Tree-structured Parzen Estimator (TPE) in proposing high-potential candidates for costly electromagnetic simulations. By integrating this method with nanophotonic design principles, we construct a reverse design model of metal structures, optimize the design of chiral structures with significant resonance peak characteristics, and advocate a chiral metasurface sensor for the highly sensitive detection of chiral molecular enantiomers. By drastically reducing the required number of full-wave electromagnetic simulations to 8000, this target-driven method achieves excellent regression performance with a validation mean squared error (MSE) of 2.1 × 10 -4 and an R 2 of 0.9969, effectively shortening the iterative design cycle to approximately ten days. Furthermore, the optimized sensor exhibits a strictly linear response in the weak handedness regime with a quantified sensitivity slope of S ≈ 0.5 , paving the way for AI-assisted chiral sensing applications in biosensing and new drug development.