Ismail Akdag, Cem Gocen, Adnan Kaya
Abstract Accurate characterization of ultra-high frequency radio frequency identification (RFID) antenna radiation patterns is critical for optimizing physical layer performance but typically entails the prohibitive costs of anechoic chambers. This study presents a novel, data-driven estimation framework that solves this inverse problem by repurposing a structured array of passive RFID tags as distributed sensors. The system operates in a semi-anechoic environment lined with microwave absorbers, where the backscatter channel is treated as a stochastic process corrupted by multipath fading and polarization mismatch. To reconstruct continuous radiation manifolds from sparse and noisy received signal strength indicator samples, a simulation-to-reality transfer learning strategy is introduced. A multi-layer perceptron regressor is trained on synthetic electromagnetic priors and subsequently applied to experimental observables to compensate for environmental losses. The framework is rigorously validated using a strict hold-out protocol across three distinct antenna architectures: a commercial Keonn reference, a custom clock-shaped prototype, and a wideband E-patch antenna. Experimental benchmarking against manufacturer datasheets and full-wave simulations demonstrates high spatial fidelity, achieving structural similarity index measures exceeding 96% and limiting mean absolute error to 0.13 dB relative to physical ground truth. This approach offers a robust, scalable, and accessible alternative to traditional vector network analyzer-based measurements for rapid prototyping and in-situ diagnostics in wireless system development.