Research on nonlinear optical noise suppression in dual-differential detection optical path based on deep learning methods

The K-Rb-²¹Ne spin-exchange relaxation free (SERF) comagnetometer is critical for ultra-high precision angular velocity measurements. However, its long-term stability is compromised by polarization noise and common-mode noise. Polarization noise arises from temperature-induced drifts in optical components, while common-mode noise is due to response mismatches between photodiodes. Existing methods only partially mitigate these issues. This paper proposes a dual differential compensation framework that integrates deep learning techniques. A Jones matrix model is employed to analyze the nonlinear coupling of these noises, and a hybrid TCN-LSTM network is used to dynamically calibrate PD response differences. Experimental results demonstrate a one-order-of-magnitude improvement in stability, with the standard deviation of the differential output reducing from 4.38 × 10⁻³ to 2.00 × 10⁻⁴ over a 2-hour period. This approach provides a robust solution for mitigating nonlinear photonic noise in high-stability quantum precision systems.