The secret key rate of continuous-variable quantum key distribution (CV-QKD) systems is limited by excess noise, primarily stemming from frequency and phase noise in transmitter and receiver lasers. This paper explores a machine learning approach using an unscented Kalman filter (UKF) for phase noise estimation in CV-QKD, comparing it to a standard reference method and an extended Kalman filter (EKF). Experimental results over a 20-km fiber link demonstrate that the UKF achieves very low excess noise, even at low pilot powers, showcasing high stability across various pilot signal-to-noise ratios. This improves robustness and simplifies CV-QKD hardware.
