The dynamics of optical soliton molecules in ultrafast lasers can reveal the intrinsic self-organized characteristics of dissipative systems. The photonic time-stretch dispersive Fourier transformation (TS-DFT) technology provides an effective method to observe the internal motion of soliton molecules real time. However, the evolution of complex soliton molecular structures has not been reconstructed from TS-DFT data satisfactorily. We train a residual convolutional neural network (RCNN) with simulated TS-DFT data and validate it using arbitrarily generated TS-DFT data to retrieve the separation and relative phase of solitons in three- and six-soliton molecules. Then, we use RCNNs to analyze the experimental TS-DFT data of three-soliton molecules in a passive mode-locked laser. The solitons can exhibit different phase evolution processes and have compound vibration frequencies simultaneously. The phase evolutions exhibit behavior consistent with single-shot autocorrelation results. Compared with autocorrelation methods, the RCNN can obtain the actual phase difference and analyze soliton molecules comprising more solitons and almost equally spaced soliton pairs. This study provides an effective method for exploring complex soliton molecule dynamics.

中文翻译：

---

使用深度神经网络重构锁模激光器中的多孤子动力学的实时光谱干扰分析

超快激光中孤子分子的动力学可以揭示耗散系统的固有自组织特征。光子时间拉伸色散傅里叶变换（TS-DFT）技术提供了一种实时观察孤子分子内部运动的有效方法。然而，尚未从TS-DFT数据令人满意地重建复杂孤子分子结构的演化。我们使用模拟的TS-DFT数据训练残差卷积神经网络（RCNN），并使用任意生成的TS-DFT数据对其进行验证，以检索三孤子和六孤子分子中孤子的分离和相对相。然后，我们使用RCNN来分析被动锁模激光器中三孤子分子的实验TS-DFT数据。孤子可以表现出不同的相变过程，并且同时具有复合振动频率。相位演化表现出与单次自相关结果一致的行为。与自相关方法相比，RCNN可以获取实际的相差并分析包含更多孤子和几乎相等间距的孤子对的孤子分子。这项研究提供了一种探索复杂孤子分子动力学的有效方法。