Deep Learning-Enabled Multi-Parameter Optimization for TWDM-PON Systems with Extended Reach

Abstract

The robustness of the physical layer greatly affects communication capacity in modern optical access networks, particularly techniques based on time-and-wavelength division multiplexed passive optical networks (TWDM-PONs). A signal may become highly vulnerable to interception, inaccurate signal reconstruction, and unstable performance due to signal degradation caused by attenuation, chromatic dispersion, and nonlinear Kerr effects. Therefore, improving transmission quality measures such as receiver sensitivity, bit error rate (BER), and Q-Factor is a prerequisite for achieving reliable and secure data delivery. In addition, deep learning (DL) is promising in improving the performance of optical access networks. This study proposes a DL-based optimizer that jointly tunes transmitter power, receiver gain, and dispersion compensation parameters. Our results demonstrate that the DL approach achieves a Q-factor above 35 dB and an average BER in the order of 10⁻¹⁷, which significantly outperform the conventional pre-forward error correction (FEC) threshold (3.8 × 10⁻³) at 80 km. Therefore, the approach significantly outperforms traditional fixed-compensation schemes which hit the FEC limit at 55 km. The optimized operating points predicted by the proposed deep neural network were independently verified using detailed OptiSystem simulations, which demonstrated strong agreement with those of the analytical model.