Fig. 1. Experimental device for wideband terahertz wireless transmission ^[83]

Fig. 2. Experimental curve of natural linewidth of Fabry-Pérot semiconductor laser using the theoretical curve to simulate the function of inverse output power^[83]

Fig. 3. Relationship between wavelength drift Δλ and normalized current (J/J_th) ^[83]

Fig. 4. Variation curves of EVM and EA output power with EA input power under 16QAM, 64QAM, and 256QAM modulation formats^[85]

Fig. 5. Typical MZM transfer curve^[86]

Fig. 6. Ideal or practical EA-aided 16QAM signal EVM and MZM relative output power curves versus MZM driving power ^[85]

Fig. 7. Variation curves of EVM and PD output power with PD input power under 16QAM, 64QAM, and 256QAM modulation formats ^[85]

Fig. 8. Schematic diagrams of different DSP equalization schemes, including traditional DSP, model-driven machine learning, and pure data-driven machine learning schemes ^[83]

Fig. 9. Schematic diagrams of NN equalizers^[92]. (a) Adaptive NN equalization; (b) CMA blind equalization based on NN; (c) proposed J-DNN equalizer

Fig. 10. Architecture of the proposed delay-tap joint DNN equalizer ^[95]

Fig. 11. BER performance vs the optical power into PD when there are 1, 2, and 3 hidden layers in DNN, respectively (solid line corresponds to the scenario when there is only one hidden layer in J-DNN)^[95]

Fig. 12. Illustration of adaptive DNN equalizer with softmax layer^[101]. (a) Sigmoid; (b) tanh; (c) ReLU

Fig. 13. LSTM channel equalization procedure^[101]. (a) Flowchart; (b) schematic diagram of the specified framework of LSTM hidden unit

Fig. 14. Relationship between BER performance and the neuron cells in hidden unit in a regular LSTM equalizer, DNN equalizer, and J-DNN equalizer, respectively^[101]

Fig. 15. GRU based model structure^[103]. (a) Detailed structure of a GRU unit; (b) structure of a dual-GRU model; (c) structure of a GRU model

Fig. 16. BER of 16QAM signal versus the input optical power^[103]. (a) Constellation diagram employing the traditional CMMA algorithm; (b) constellation diagram with dual-GRU

Fig. 17. Neural framework of our proposed fully complex valued DNN equalizer^[101]

Fig. 18. BER performance vs the optical power into PD for 45 Gbaud PAM-4 signal wireless transmission by employing 45-tap CMMA equalizer combined with 301-tap DD-LMS, nonlinear RVNN and CVNN equalizers, respectively^[101]

Fig. 19. Detailed DSP blocks at the Tx- and Rx-side^[101]

Fig. 20. DSP steps and signal constellation diagram (insets: consellation diagrams after down-conversion, FOE, CPR, and RVNN equalization)^[101]

Fig. 21. Scheme of pre-processing before entering NN classifier^[106]

Fig. 22. Principle of random under-sampling

Fig. 23. Principle of random oversampling and target labels of different datasets^[106]. (a) Original dataset; (b) random oversampling dataset; (c) balance dataset

Fig. 24. Real-valued NN classifier^[106]. (a) Schematic structure; (b) complex-valued NN classifier with a cross entropy loss function

Fig. 25. BER performance vs the optical power into PD for 10 Gbaud THz PS-64QAM signal wireless transmission by employing 21-tap CMMA equalizer combined with 223-tap DD-LMS, 201-2^nd tap Volterra equalizer, nonlinear RVNN, and CVNN classifiers, respectively^[106]

Fig. 26. CVNN with Logit adjustments^[107]

Fig. 27. BER of PS-16QAM when receiving power changes from -5 dBm to 1 dBm^[107]