Machine Learning‐Assisted Design and Performance Prediction of a Compact Dual‐Band Polarization‐Insensitive THz Metamaterial Absorber for Skin‐Cancer‐Related Refractive‐Index Sensing
Paper details
Section titled “Paper details”| Field | Value |
|---|---|
| Journal | Advanced Electronic Materials |
| Published | 2026-06-01 |
| Authors | Md. Murad Kabir Nipun, S. M. Rifatur Rana, Md Golam Morshed, Imrose Jahan, F A Sabbir Ahamed |
| DOI | 10.1002/aelm.70449 |
| Citations | 2 |
The simulated Quad-Arrow Split Ring Resonator absorber reached 99.96% absorption at 4.8 THz and 99.91% at 3.3 THz. Its unit cell measured 0.24 λ × 0.24 λ × 0.03 λ at the first resonance. Both bands shifted downward as analyte refractive index increased from 1.35 to 1.40. These results apply to a homogeneous 1 µm dielectric layer, not clinical tissue or a fabricated device.
Geometry and full-wave method
Section titled “Geometry and full-wave method”The QASRR combines a diamond-shaped outer metallic frame with a central split circular ring and four inward arrow elements. Fourfold symmetry supports polarization-insensitive response at normal incidence.
The split ring alone produced weak absorption without a dual-band response. Adding the four arrows produced a band near 4.8 THz, while the outer frame added the mode near 3.3 THz. The selected inward-triangle displacement was u = 0.3, and the optimized substrate thickness was 2.5 µm.
The stack uses a 2.5 µm polyimide spacer with εr = 3.5 and tan δ = 0.0027. The patterned resonator and ground plane each use 0.1 µm silver. Silver conductivity was set to 6.3012 × 10^7 S/m. Its calculated skin depth at 3.3 THz was 0.035 µm, so the modeled silver was 2.87 times the skin depth.
CST Microwave Studio 2019 used its frequency-domain solver on one periodic unit cell. The solver covered 0.1 to 10 THz, while the reported absorber response was examined over 2.5 to 5.5 THz. No fabricated prototype or physical spectrum was reported.
Dual-band response and resonance mechanism
Section titled “Dual-band response and resonance mechanism”Absorption was calculated from A = 1 - R - T using the simulated reflection and transmission coefficients. The continuous silver ground suppressed transmission, leaving reflection minima at 3.3 and 4.8 THz.
Effective permittivity, permeability, and impedance were retrieved from S11 and S21 through the Nicolson-Ross-Weir method. The retrieved normalized impedance was 0.987 + j0.076 at 3.3 THz and 0.963 - j0.058 at 4.8 THz. Both values are close to the free-space matching condition of Re(Z) = 1 and Im(Z) = 0.
At 3.3 THz, electric energy occupied the outer frame edges and ring sides, while magnetic energy concentrated around the ring and side arrows. At 4.8 THz, the electric field localized near the central vertical axis and upper and lower arrows. The magnetic field shifted toward the left and right arrow regions, indicating a different current loop.
Surface currents on the resonator and ground flowed in opposite directions at both frequencies. This anti-parallel flow supports magnetic coupling through the polyimide spacer.
An ADS circuit with two parallel series RLC branches reproduced resonances near 3.27 and 4.86 THz. The upper branch used L1 = 17.15 pH, R1 = 49 Ω, and C1 = 0.138 fF. The lower branch used L2 = 31.55 pH, R2 = 50.2 Ω, and C2 = 0.034 fF. The circuit and CST curves agreed near resonance but differed away from the peaks.
Polarization rotation and incidence angle were each tested from 0° to 60°. Both peaks remained identifiable, with minor amplitude changes and negligible frequency displacement.
Refractive-index sensing under a homogeneous analyte
Section titled “Refractive-index sensing under a homogeneous analyte”The analyte was a homogeneous dielectric layer fixed at 1 µm thickness. Its refractive index was swept from 1.35 to 1.40. The authors used n = 1.360 for normal skin cells and n = 1.380 for cancer-related skin cells. Both resonances moved toward lower frequency as refractive index increased.
The following values apply to the 1 µm homogeneous analyte model.
| Metric | 3.3 THz mode | 4.8 THz mode |
|---|---|---|
| Peak absorption | 99.91% | 99.96% |
| Sensitivity | 2950 GHz/RIU | 4846 GHz/RIU |
| Figure of merit | 32.13 | 21.5 |
| Q-factor | 14.27 | 12.8 |
The second mode gave the larger frequency-shift sensitivity, while the first mode retained better spectral quality. The ML model was not trained to extract sensitivity or optimize the sensing response. All sensing metrics came from CST resonance-shift sweeps.
Real tissue has a complex refractive index, and its response can depend on hydration, water distribution, composition, preparation, and measurement conditions. The paper therefore limits the result to controlled dielectric-contrast sensing.
Comparison with recent absorbers
Section titled “Comparison with recent absorbers”The paper’s comparison table places the dual-band design against recent THz absorbers. The proposed first-mode sensitivity of 2950 GHz/RIU exceeds 300 GHz/RIU in Ref. [5] and 1694.7 GHz/RIU in Ref. [57]. The second value of 4846 GHz/RIU also exceeds 3000 GHz/RIU in Ref. [59].
The first-mode figure of merit of 32.13 exceeds 22.045 in Ref. [57] and the listed values in Ref. [58]. The Q-factors do not lead the comparison. Ref. [5] reports 61.75, and Ref. [57] reports 78.9. Prior work therefore supports competitive sensing, but not leadership on every metric.
Peak absorption is comparable with other near-unity results. The proposed combination also includes compact dual-band operation, angular tolerance, circuit agreement, and surrogate prediction.
StackNet prediction of simulated absorption
Section titled “StackNet prediction of simulated absorption”The ML dataset contains 223,573 simulated absorption samples from 216 combinations of substrate, resonator, and ground-plane thickness. Substrate thickness ranged from 2.1 to 3.1 µm. Resonator and ground thicknesses each ranged from 0.1 to 1.1 µm. Spectra covered 2.5 to 5.5 THz, with frequency and three thicknesses as raw inputs.
TC-40 used 60% of the samples for training, TC-50 used 50%, and TC-60 used 40%. Thirteen physics-based features expanded the input to 17 dimensions. These included frequency squared, frequency times substrate thickness, frequency divided by substrate thickness, and the square root of resonator thickness.
StackNet combined KNN, Random Forest, Extra Trees, and XGBoost as base learners. A Ridge meta-learner used out-of-fold predictions to avoid information leakage.
The abstract and conclusion report an average R² of 0.9986 across the three cases. Section 7.1 instead lists StackNet R² values of 1.0000 for TC-40 and 0.9999 for TC-50 and TC-60. The paper does not reconcile these figures. Extra Trees alone reached R² = 0.9999 in all three cases, so the reported ensemble gain is small.
The Ridge weights were 0.713 for Extra Trees, 0.208 for KNN, 0.059 for Random Forest, and 0.021 for XGBoost. SHAP ranked frequency squared, frequency, frequency times substrate thickness, frequency divided by substrate thickness, and the square root of resonator thickness as the five most influential features.
All four models produced negative R² values from -2.021 to -1.853 when trained on 3.28 to 5.03 THz and tested outside that range. At 5% input noise, Extra Trees and KNN retained R² values of 0.9778 and 0.9809.
Proposed fabrication and validation limits
Section titled “Proposed fabrication and validation limits”The authors outline a route for the multilayer stack. Thermal evaporation or sputtering can form the 0.1 µm ground silver. Spin coating and curing can form the 2.5 µm polyimide spacer. Photolithography or electron-beam lithography defines the QASRR before 0.1 µm silver deposition and lift-off.
They propose THz time-domain spectroscopy in dry air or nitrogen, with reflection measured against a metallic mirror. No prototype, fabrication tolerance study, or THz-TDS result was reported.
Expected deviations include lithographic tolerance, edge rounding, silver roughness, thin-film conductivity variation, and nonuniform polyimide thickness. A Ti or Cr adhesion layer may also be needed between silver and polyimide.
Sensing adds uncertainty from analyte thickness, incomplete contact with the high-field region, and complex tissue dielectric properties. The sensing result therefore remains a controlled simulation until a fabricated device is measured with controlled dielectrics and realistic biological samples.
Engineering implications
Section titled “Engineering implications”The two modes support different instrument choices. The 3.3 THz mode offers the better linewidth-related performance, while the 4.8 THz mode gives the larger resonance displacement.
The ML models are interpolation tools for the simulated parameter space. They should not guide operation outside the 2.5 to 5.5 THz frequency range or the sampled thickness range. The next design gate is a fabricated polyimide and silver stack whose resonance shifts are measured with a controlled dielectric and then repeated with realistic tissue.
View Original Abstract
ABSTRACT The proposed metamaterial absorber is built on a compact Quad‐Arrow Split Ring Resonator (QASRR) consisting of a diamond‐shaped structure with a centrally loaded split circular ring with four arrow‐shaped extensions. The absorber exhibits polarization‐insensitive dual‐band absorption at 3.3 and 4.8 terahertz (THz), with peak absorptivities of 99.91% and 99.96%, respectively. Its normalized size is 0.24λ × 0.24λ × 0.03λ at the first resonance. Using representative refractive indices for normal and cancer‐related skin‐cell models ( n = 1.360 and n = 1.380), the absorber resolves a small refractive‐index contrast under a controlled homogeneous analyte model, with sensitivities of 2950 GHz/RIU and 4846 gigahertz per refractive index unit (GHz/RIU). The corresponding figure of merit (FOM) values are 32.13 and 21.5, with Quality‐factors of 14.27 and 12.8. Field distributions, surface currents, substrate‐dependent parametric studies, and equivalent circuit modeling are used to explain the dual‐resonance mechanism and validate the results. The main contribution is a unified framework that links compact QASRR‐based dual‐band absorption, skin‐cancer‐related refractive‐index sensing, field/current and circuit‐based electromagnetic mechanism interpretation, and StackNet‐assisted response prediction. StackNet is developed using physics‐informed engineered features and compared with k‐nearest neighbor (KNN), random forest (RF), extra trees (ET), and extreme gradient boosting (XGBoost) models. It achieves an average coefficient of determination (R 2 ) of 0.9986 across TC‐50, TC‐60, and TC‐40 datasets, indicating reliable prediction of the simulated absorber response.
Advanced Thermal Simulation Service Diamond Heat Spreader Performance Analysis
View Specs →
Free Standing Boron Doped Diamond Pcd Film
View Specs →
Electronic Grade Intrinsic Diamond Epitaxial Wafer H Terminated Superior 100 Mosfet Ready Substrate
View Specs →Original source
Section titled “Original source”References
Section titled “References”- 2023 - Terahertz Metamaterials for Biosensing Applications: A Review [Crossref]
- 2024 - Review on the Terahertz Metasensor: From Featureless Refractive Index Sensing to Molecular Identification [Crossref]
- 2015 - Experimental Demonstration of Ultrasensitive Sensing With Terahertz Metamaterial Absorbers: A Comparison With the Metasurfaces [Crossref]
- 2023 - High‐Performance Terahertz Refractive Index Sensor for Cancer Cells Detection [Crossref]
- 2025 - Skin Cancer Detection Using Terahertz Metamaterial Absorber and Machine Learning [Crossref]
- 2023 - Analysis of a Thin Tunable Silicon‐Based Metamaterial Absorber for Sensing Applications [Crossref]
- 2023 - Design of Graphene‐Constituted Frequency Selective Surface ‐Based Perfect Absorbers in THz Regime—A Circuit Analysis Approach [Crossref]
- 2025 - Z‐Slotted Silicon Dioxide Based Tunable Polarization Converter Design Using Graphene‐VO2 Composite for Detection of Cancer [Crossref]
- 2025 - Tunable THz Sensing for Early Detection of Skin Cancer by Deep Learning Enabled Image Reconstruction [Crossref]
- 2025 - Acoustic Moiré Flat Bands in Twisted Heterobilayer Metasurface [Crossref]