Deep-Learning Empowered Customized Chiral Metasurface for Calibration-Free Biosensing

Area of Science:

• Photonics and Materials Science
• Artificial Intelligence in Scientific Discovery

Background:

• Metasurfaces are 2D metamaterials enabling advanced light-matter interactions.
• Traditional metasurface design is slow and iterative, limiting exploration of intelligent design strategies.
• Data-driven approaches for metasurface design are underexplored, especially for chiral applications.

Purpose of the Study:

• To develop a novel data iterative strategy using deep learning and global optimization for customized chiral metasurface design.
• To enable precise, label-free identification of chiral molecules.
• To overcome limitations of traditional simulation-based data generation in exploring the full design space.

Main Methods:

• A deep learning-based data iterative strategy coupled with a global optimization network was employed.
• A novel data generation strategy was developed to encompass the entire design space, surpassing conventional simulation limitations.
• The methodology was applied to design chiral metasurfaces for molecular identification.

Main Results:

• A 21-fold increase in chiral structures with a desired circular dichroism (CD) response (>0.6) was achieved.
• The method demonstrated superior dataset quality and design space exploration compared to traditional approaches.
• A validated monolayer structure showed advanced sensing capabilities for enantiomer-specific analysis of bio-samples.

Conclusions:

• Data-driven schemes offer superior capabilities for photonic design.
• Chiral metasurface platforms show significant potential for calibration-free biosensing.
• The proposed approach accelerates development for molecular detection and spectroscopic imaging.