Electromagnetically Consistent Evaluation of Federated RIS Channel Estimators: Hidden Idealizations and a Full-Wave Virtual Testbed
IEEE ACCESS, cilt.14, ss.167306-167330, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3748055
- Dergi Adı: IEEE ACCESS
- Derginin Tarandığı İndeksler: Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.167306-167330
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Erzincan Binali Yıldırım Üniversitesi Adresli: Evet
Özet
Learned reconfigurable intelligent surface (RIS) channel estimators are typically evaluated on an idealized diagonal phase-shift surface, with noise added to the channel and observations one nominal slot apart. This paper removes these idealizations and tests which conclusions persist. An electromagnetically consistent generator (a varactor-loaded cell calibrated to full-wave simulation, multiple-scattering mutual coupling, spherical wavefronts) is driven by an M +1 pilot sweep under non-Markovian Jakes dynamics; a federated long short-term memory estimator is compared with classical filters, including a deployable Kalman filter. Two idealizations systematically favor the learned estimator. Referring the cell reflection to the wave impedance instead of the element self-impedance models the passive surface as active and inflates the least-squares error 9- to 87-fold across five cell families under a dipole kernel, against about threefold for the learned estimator. Spacing observations by one slot overstates temporal correlation; retrained for the corrected interval, the recurrent estimator has 57% higher NMSE than its nominal-slot counterpart. Under the corrected models the deployable Kalman filter leads except at small normalized Doppler. Despite a 3.6- to 6.6-fold column-NMSE lead over least squares, the learned estimator delivers less field at the user in all 60 combinations of RIS configuration, beamforming or RIS-control strategy and training seed: its residual survives the matched-filter projection and is mostly amplitude error, so error-phase coherence predicts the delivered field better than NMSE. A virtual over-the-air testbed, which adds hardware-impairment models to an 84-port full-wave network with multipath, confirms the passivity audit and the ordering of simulation-trained estimators.