Datasets for Model Training, Validation, and Testing, Together with Spectral Measurements and Optical Characterization of Eight Devices

A dataset comprising 640 spectra of thermally stabilized DKSs is partitioned into training and validation sets in a 4:1 ratio. The ground-truth soliton number labels required for semi-supervised model training are first obtained through FFT-GA-based spectral analysis. These labeled datasets are then used to train a hybrid CNN-TMHA model, enabling efficient extraction of representative local and global spectral features. Finally, a testing dataset of 230 real-time acquired soliton spectra is fed into the trained model to enable rapid and accurate classification of distinct DKS states.<br>Based on the spectral measurements and optical characterization results of the eight devices, the following observations can be made:(1) As the coupling gap increases, the resonance modes progressively transition from the over-coupled regime to the critically coupled regime, and eventually to the under-coupled regime. (2) The overall symmetry of the dispersion curves about the central mode confirms that the dispersion landscape is only weakly perturbed.

本数据集包含640份热稳定耗散克尔孤子（Dissipative Kerr Solitons, DKS）光谱，并以4:1的比例划分为训练集与验证集。半监督模型训练所需的真实孤子数量标签，首先通过基于快速傅里叶变换-遗传算法（Fast Fourier Transform-Genetic Algorithm, FFT-GA）的光谱分析方法获取。随后利用该标注数据集训练混合CNN-TMHA模型，实现代表性局部与全局光谱特征的高效提取。最后，将包含230份实时采集孤子光谱的测试集输入至训练完成的模型，实现不同DKS状态的快速精准分类。<br>基于8台器件的光谱测量与光学表征结果，可得到如下观测结论：(1) 随着耦合间隙增大，谐振模式将依次从过耦合态过渡到临界耦合态，最终进入欠耦合态。(2) 色散曲线围绕中心模式的整体对称性表明，色散特性仅受到微弱扰动。