|
|
 |
| |
|
|
|
Hybrid CNN-Transformer for Multi-Band THz MIMO Antenna Optimization: 6G Deep Learning |
|
|
|
PP: 1279-1289 |
|
|
doi:10.18576/amis/200513
|
|
|
|
Author(s) |
|
|
|
Hamza A. Mashagba,
Hamza Abu Owida,
Manal Mizher,
Asokan Vasudevan,
Azlan B. Abd Aziz,
Mohammad Faleh Ahmmad Hunitie,
Eyad Dhaher Hassan Megdadi,
Suleiman Ibrahim Mohammad,
|
|
|
|
Abstract |
|
|
| Designing THz MIMO antennas for 6G wireless communications entails significant computation costs due to complex interactions between geometric and electromagnetic parameters. Traditional methods rely heavily on EM simulations, which take an enormous amount of time and computational power. In this paper, a new hybrid deep learning model based on CNNs and Transformer models was proposed that allows the simultaneous prediction of multiple performance metrics of graphene THz MIMO antennas. Our architecture considers twelve geometric and material parameters to predict six metrics of interest, namely S11, Gain, Isolation, Bandwidth, Envelope Correlation Coefficient (ECC), and Diversity Gain (DG). For training our deep learning algorithm, 756 different antenna designs over 1–6 THz were created using CST Microwave Studio software. The proposed hybrid architecture provides highly accurate predictions with R2 values of 98.73%, 97.91%, and 98.45% for Gain, Isolation, and Bandwidth, respectively, outperforming conventional machine learning algorithms like Random Forest (R2 = 94.23%), XGBoost (R2 = 95.67%), and Deep Learning models (e.g., CNNs with R2 = 96.84% and Transformers with R2 = 97.12%). With our proposed architecture, the simulation process has been accelerated by 1350 times; from 45 minutes to only 2 seconds, while the accuracy has stayed within the 2.3% MAPE. Ablation experiments show that both CNN’s local feature extraction and Transformer’s global dependency modeling are required to capture the necessary interactions.
Keywords: |
|
|
|
|
|
|
|
