|
|
 |
| |
|
|
|
Hybrid CNN-Transformer Deep Learning Architecture for Terahertz Metamaterial Absorption Prediction |
|
|
|
PP: 539-548 |
|
|
doi:10.18576/amis/200219
|
|
|
|
Author(s) |
|
|
|
Hamza A. Mashagba,
Hamza Abu Owida,
Suhaila Abuowaid,
Suleiman Ibrahim Mohammad,
Azlan B. Abd Aziz,
Manal Mizher,
Asokan Vasudevan,
Mardeni Bin Roslee,
|
|
|
|
Abstract |
|
|
| Terahertz (THz) metamaterial absorbers are key to many of the emerging technologies including 5G/6G communication systems, medical imaging techniques, as well as security scanning systems. Nevertheless, electromagnetic simulation based design methods have been shown to be severely limited by computation requirements which can take from 5–30 min to evaluate and thus limit how much the design space can be explored. Therefore this research has developed a new hybrid deep learning methodology for THz metamaterial absorber design that uses a combination of CNN’s (Convolution Neural Networks) to extract hierarchical geometric features from images of metamaterial designs; and, also utilizes transformer encoder blocks using multi head attention mechanism to capture frequency dependent absorption behaviors. We conduct extensive experimental evaluation on 9,018 simulation- generated samples spanning 5–11 THz frequencies, with systematic variations in patch width (20–80 μm) and dielectric thickness (5–20 μm), using rigorous 70/15/15 train/validation/test splitting to ensure representative distributions. Our hybrid CNN-Transformer achieves state-of-the-art performance with R2 = 0.9995, MAE = 0.0098, and RMSE = 0.0135, representing statistically significant improvements (p < 0.001, Cohen’s d = 2.34) over Random Forest (R2 = 0.9985), XGBoost (R2 = 0.9981), pure CNN (R2 = 0.9963), and LSTM (R2 = 0.9951) baselines. Comprehensive ablation studies across eight architectural configurations reveal that CNNs contribute 82.6% of predictive performance through spatial feature extraction, while Transformers add 17.4% via long-range dependency modeling. Bootstrap confidence intervals [R2: 0.9993–0.9997, MAE: 0.0095–0.0101, RMSE: 0.0131–0.0139] demonstrate high reliability, while five-fold cross-validation reveals excellent stability (coefficient of variation < 0.5%).The top three predictors of antenna performance are frequency (43%); patch width (27%), and dielectric thickness (20%) identified through permutation-based feature importance. The computational efficiency study found that each model is able to perform an inference for a given example in approximately 4.2 ms; this represents a 239× increase in speed relative to CST Microwave Studio simulation times; therefore, our system enables real-time design exploration of antennas via simulation which was previously not possible due to the slow computation times associated with traditional EM solvers.
Keywords: |
|
|
|
|
|
|
|
