End-to-End Optimization of Constellation Shaping Robust to Semiconductor Laser Noise and Nonlinearity in Fiber-THz System

We propose a novel end-to-end (E2E) fiber-terahertz (THz) integrated system based on a data-driven mutual information neural estimator (MINE). This system utilizes the MINE neural networks for Tx-NNs optimization, enabling constellation shaping, channel sampling without known channel models, and a differentiable carrier phase estimation for resisting phase noise. Our approach follows state-of-the-art bitwise autoencoders (AE), which requires a differentiable implementation of all operations between transmitter and receiver, including the DSP algorithms. Firstly, to address the demand for higher spectral efficiencies, we employ the E2E-AE framework for constellation shaping in the presence of laser noises and nonlinearity. Subsequently, we rigorously examine the characterization of lasers with white and colored frequency modulation noise and investigate their impact on MI performance using simulation and experiment when the carrier phase recovery algorithm is present and absent. Finally, We experimentally evaluate our proposed E2E framework within a dual-polarized fiber-THz integrated system at 320 GHz, considering linewidth, flicker ($1/f$) noise, and resonance peaks. By joint optimization constellation shaping and compensation for the phase noise and nonlinearity, we achieve a robust and pilot-free modulation scheme improving the performances by up to 0.15-bit and 1.21-dB gains in MI and SNR compared to Grid-QAM constellations. Additionally, under the constraints of laser phase noise and nonlinearity, we achieve a transmission speed of 230 Gbit/s in the E2E-AE fiber-THz system for the first time. This solution advances the E2E-AE framework by incorporating differentiable functions tailored for practical scenarios.