Deep learning-assisted sustainable design of a multi-substrate 2.4 GHz wearable IoT antenna

Wearable antennas are essential components of modern Internet of Things (IoT) systems, enabling wireless communication for body-centric, healthcare, environmental-monitoring, and smart-textile applications. Deep learning provides an effective means of accelerating antenna design and optimization under complex material and operating conditions. This paper presents a deep learning-assisted and sustainability-oriented framework for designing a compact 2.4 GHz wearable IoT antenna implemented on three representative substrates: FR4, Rogers laminate, and felt-based textile. A unified slotted microstrip patch topology is adopted to enable a consistent comparison between rigid and flexible antenna realizations while highlighting trade-offs among impedance matching, bandwidth, mechanical flexibility, and material characteristics. A multi-substrate full-wave electromagnetic dataset containing approximately $$1.09\times {10}^{5}$$ samples is generated in CST Microwave Studio by varying the principal geometrical parameters, substrate descriptors, and operating frequency over the 2.0–3.0 GHz range. A fully connected deep neural network surrogate is trained to predict the reflection coefficient from antenna geometry, substrate properties, and operating frequency, achieving an RMSE of 0.8500 dB and $${R}^{2}=0.9553$$ on the held-out sample-level test partition. A separate geometry-level holdout evaluation is also conducted using six antenna configurations whose complete frequency responses are excluded from training and validation. The optimized antenna designs are fabricated on the three substrates and experimentally evaluated through calibrated reflection-coefficient measurements. For the felt-based textile prototype, additional measured reflection-coefficient responses under on-arm bending and progressive surface-wetting conditions demonstrate stable impedance matching around the 2.4 GHz ISM band. Radiation patterns, realized gain, radiation efficiency, on-body behavior, and specific absorption rate are evaluated using full-wave simulations only; no experimental over-the-air radiation validation is claimed. The simulations predict broadside radiation behavior and simulated SAR values below the applicable international safety limits. An equivalent parallel RLC model is also introduced to support the physical interpretation of substrate-dependent impedance behavior. The results indicate that the proposed surrogate-assisted framework enables rapid, substrate-aware antenna design and reduces dependence on repeated full-wave simulations. The felt-based implementation provides a practical compromise among simulated electromagnetic performance, flexibility, reduced substrate mass, and wearable suitability.