International Journal of Sustainable Building Technology and Urban Development ISSN:2093-761X(Print) 2093-7628(Online)

The development of sustainable smart buildings and intelligent urban infrastructure increasingly depends on energy-efficient high-speed wireless communication and sensing devices for structural monitoring, building automation, and infrastructure management. Structural health monitoring (SHM) devices require reliable techniques to detect early-stage damage, identify defect locations, analyse structural stability, and ensure safe and reliable operations of critical infrastructure. Terahertz (THz) technology has emerged as a effective solution than other lower frequencies because THz waves are non-ionizing, safe for human environments, capable of penetrating many non-metallic materials, and capable of achieving high-resolution image and material characterization. However, the design and optimization of Terahertz (THz) Multiple-Input Multiple-Output (MIMO) antennas for sixth-generation (6G) Internet of Things (IoT) devices are constrained by the high computational cost of full-wave electromagnetic simulations. To address this challenge, this study presents a high-fidelity population-centric deep learning framework, termed Pocaii-DNN, for rapid and accurate electromagnetic modeling of THz MIMO antennas in sustainable building and urban sensing applications. The proposed framework employs four parallel Adaptive Intelligence Units (AIUs) to independently process frequency, parametric analysis of patch geometry, slot geometry, and feed position parameters, followed by a Cooperative Fusion mechanism to capture nonlinear interactions among heterogeneous antenna design variables. The model is trained as a regression-based electromagnetic emulator to predict the antenna reflection coefficient (S11) directly from design parameters. Experimental results demonstrate solver-grade accuracy, achieving a mean absolute error (MAE) of 0.051346 dB, a mean squared error (MSE) of 0.003735 dB², and a coefficient of determination (R²) of 0.997648 on unseen test datasets. Comparative analysis with Random Forest and Decision Tree models confirms superior prediction performance and generalization capability. Furthermore, the trained framework generates S11 predictions within milliseconds, significantly reducing computational time and energy consumption compared with conventional HFSS simulations. The optimized absorber-assisted THz MIMO antenna operates over 2.20 THz to 5.50 THz with a wide bandwidth of 3.30 THz and exhibits excellent MIMO performance, resulting an ECC of 0.00814, a DG of 9.9997 dB, a TARC of −14.35 dB, and a CCL of 0.8386 bits/s/Hz providing improved impedance matching, bandwidth response, noise performance, and quality factor, ensuring stable operation at terahertz frequencies. The proposed approach enables efficient virtual prototyping and optimization of THz MIMO antennas for sustainable smart buildings, integrated sensing and communication (ISAC) applications, enabling crack, void, moisture, and material degradation detection in structural health monitoring systems., and intelligent urban infrastructure system.