Data Driven Inverse Design of Optical Metamaterials

 

The unique optical properties of metamaterials hold the potential to revolutionize fields like displays, imaging, and sensing. However, unlocking this potential requires efficient inverse design algorithms capable of accurately predicting the complex structures that yield desired optical responses. Traditional methods struggle with the vast design space and computational demands of complex metamaterials.

To address these challenges, we introduce prior knowledge into deep learning (DL) frameworks for accurate and scalable inverse design of optical metamaterials. This carefully selected prior knowledge reduces data requirements and enhances model robustness. We integrate the prior knowledge as (a) governing physics equations for simplified structures and (b) learned parameters from DL models trained on simpler designs.

Focusing on plasmonic metamaterials with metal gratings on a dielectric film, we consider two designs: (a) simple periodic metal blocks and (b) graded metal blocks with varying widths. Our developed DL models achieve impressive performance: 97.4% accuracy for predicting design parameters in simple blocks, 86% accuracy for diffraction efficiencies in graded blocks, and 97% accuracy for spatial electric field distributions.

Furthermore, we develop an auto-regressive transformer-based DL model for the inverse design of adaptive metamaterials with dynamic components. This model captures the complex relationships between dynamic optical responses and the metamaterial's design, including both static and time-varying elements. It achieves 98.5% accuracy for material prediction, 82.5% for grating profile design, and 95% accuracy for dynamic thickness prediction.

We also apply our DL expertise to a critical real-world challenge: the inverse prediction of blood constituent concentrations for sample integrity in medical diagnostics. To overcome the limitations of obtaining extensive real blood samples, we develop a physics-based forward model that simulates the spectral response of blood samples under various conditions. This model generates the data needed to train a robust DL model for accurately predicting hemoglobin and bilirubin concentrations. Our DL model achieves an accuracy of up to 99% for hemoglobin and bilirubin concentration prediction, ensuring reliable sample integrity analysis for medical diagnostics.