Machine learning-assisted design and FEM simulation of a multi-material metasurface terahertz refractive index sensor
Abstract
In this paper, the design, numerical simulation, and optimization of label-free terahertz refractive index sensor using periodic metasurface are presented. The sensor is made of an array of 8 unit-cells on a silicon dioxide (SiO₂) substrate. Elliptical resonators with gold (Au), silver (Ag), MXene and phosphorene are placed on top of square coupling layers coated with copper (Cu).The finite element method (FEM) is used in the COMSOL Multiphysics simulation software to calculate the electromagnetic responses. The simulations are based on the Floquet–Bloch periodic boundary conditions and extended Drude–Lorentz dispersion model, which describes the material behavior in the terahertz regime.The four structures are tested for their properties over the range of RIU values of the analyte from 1.3333 to 1.4833. The performance of the detector is judged by sensitivity, figure of merit (FOM), quality factor, detection limit, signal to noise ratio and dynamic range. The fourth performance configuration provides the best performance (1 THz/RIU peak sensitivity, and 26.316 RIU −1 FOM). The performance gradually increases from one configuration to the other, and the achieved results of the evaluated metrics are around 10–15% higher.The machine learning regression model is used to predict the sensor response as a function of angle of incidence (0° to 80°) and square resonator lateral dimension (1.3 to 5.3 μm). For all test cases, the model's R 2 values are greater than 0.998 and RMSE values are less than 0.007, indicating that the response behavior can be quickly predicted without full-wave simulation. The metasurface is intended for label-free bulk refractive-index sensing and tissue dielectric characterization—applications including surgical-margin assessment and ex vivo tissue discrimination in medical diagnostics, together with food-safety and environmental-monitoring tasks that rely on bulk dielectric contrast rather than single-molecule recognition.
