Neural Network Embedded With Maximum Likelihood Phase Retrieval for Photonics-Aided Wireless Transmission System at 0.33–0.52 THz

Photonics-assisted terahertz (THz) wireless communication systems have emerged as a promising solution for future THz-over-fiber (ToF) networks, offering both abundant bandwidth and seamless integration with optical infrastructure. However, existing THz systems typically operate within limited bandwidths (i.e., <0.1 THz), restricting their adaptability for diverse THz applications. Furthermore, laser phase noise and system nonlinearities significantly impair signal recovery at the receiver, while conventional approaches employing separate phase compensation and nonlinear equalization suffer from both high computational complexity and performance degradation. In this article, we present a novel low-complexity neural network architecture embedded with maximum likelihood phase retrieval (NN-eMLPR) for photonics-aided wireless transmission across the 0.33–0.52-THz band. Our solution integrates a maximum likelihood phase recovery layer directly within the neural network (NN) structure, enabling simultaneous equalization and phase noise compensation. Experimental results demonstrate successful 30-GBaud signal transmission throughout the 0.33–0.52-THz range. The NN-eMLPR achieves a 44.6% reduction in multiply-accumulate (MACC) operations compared to traditional phase recovery methods with NN while maintaining superior bit-error-rate (BER) performance, offering both computational efficiency and enhanced signal integrity across this ultrawide bandwidth.