Label-Free Environment Adaptation for Wireless Receivers via Physics-Informed Constraints

Abstract

Adapting deep-learning receivers after deployment is challenging because ground-truth channel state and transmitted payload symbols are unavailable, while sparse pilots alone provide insufficient supervision. The orthogonal frequency-division multiplexing waveform, however, embeds physical structure that can provide label-free supervision through physics-informed constraints. Here, we advance Physics-Embedded Inverse Learning (PEIL) toward deployment in two phases. Phase I develops Improved PEIL, a compact receiver with 0.093M trainable parameters that coordinates carrier frequency offset estimation and channel state information refinement through symbol-recovery supervision. In Phase II, Physics-Embedded Label-free Adaptation (PELA) adapts the same receiver using pilot reconstruction together with cyclic-prefix phase consistency and delay–symbol smoothness of the channel impulse response tail. Trained through 16-QAM symbol recovery, Improved PEIL generalizes without retraining to unseen QPSK and 64-QAM payloads. Under the EVA-to-ETU transition at 35 dB, PELA reduces symbol error rate from 0.1501 to 0.0258 and block error rate from 0.7106 to 0.0591 relative to static Improved PEIL. Across the evaluated mobility shifts, PELA becomes beneficial when deployment conditions differ substantially from the training environment. These results show that protocol-native pilots and waveform physics can supply the missing supervision for selective post-deployment adaptation.