Photonic-assisted terahertz (THz) systems, owing to their ultra-wide modulation bandwidth, are widely regarded as an ideal solution for emerging 6G applications. Meanwhile, multiple-input multiple-output (MIMO) technology leverages multi-antenna arrays to form multidimensional independent channels, enhancing link quality and system capacity, and is considered a key candidate for high-performance THz communications. To further unleash the potential of MIMO, advanced digital signal processing (DSP) techniques and their scalability in MIMO architectures have become active research topics. However, conventional DSP shows limited performance under complex channel conditions, whereas supervised neural-network approaches generally incur high computational complexity. To address these challenges, we propose an unsupervised deep neural-network equalizer based on the constant-modulus algorithm (CMA) for blind equalization of QAM signals, and we verify its favorable scalability in MIMO systems. In a dual-polarization 2×2 MIMO transmission experiment, 50 Gbps QPSK transmission was successfully achieved at 128.75 GHz, with a bit error rate (BER) of 6.3×10 -3 . Experimental results show that the proposed Blind DNN equalizer requires no training data and exhibits a stronger resistance to overfitting than supervised DNNs. In single-channel transmission, it achieves a 1.5 dB signal-to-noise ratio (SNR) gain over T/2-CMA; when combined with maximum-ratio combining (MRC), the system attains a 2.8 dB diversity gain. Furthermore, the proposed MIMO-Blind DNN-MRC integrates the combining process into the equalizer, significantly reducing the complexity while maintaining performance, requiring only a single DSP path. This study provides a feasible pathway for integrating traditional DSP with data-driven neural networks, and offers empirical evidence for validating the scalability of deep learning in THz MIMO systems.
Pan et al. (Wed,) studied this question.