脊波导的单模特性对于高度集成的光电元件和片上激光器有极其重要的作用，但传统时域有限差分算法复杂且计算资源消耗较大。提出并验证了一种基于深度学习的绝缘体上硅（silicon-on-insulator, SOI）脊波导单模特性快速识别方法。通过设计一种结合贝叶斯算法优化后的人工神经网络的脊波导模态分类方法，实现了高精度脊波导模态的快速分类。通过学习脊波导几何结构与单模特性的映射关系，使网络可以预测未知脊波导几何参数的单模特性。结果显示，该模型仅用252个训练数据即可在测试集上达到99%的分类精度，单数据点的分类时间由传统时域有限差分算法52 s降为5 ms，计算效率提升了约1 000倍。同时，该模型具有较好的泛化能力。此算法有望缩短光波导的设计周期，这对于大规模光子芯片的设计和制备有一定的支持作用。

The single-mode of ridge waveguides is important to integrated optoelectronic components and on-chip lasers. The traditional finite-difference time-domain method consumes a lot of computing resources to design the mode. A deep learning framework was proposed for rapid identification of single-mode regimes in silicon-on-insulator (SOI) ridge waveguides. A Bayesian-optimized artificial neural network was applied to modal classification. By designing a ridge waveguide mode classification method that combines an optimized artificial neural network with Bayesian algorithms, high-precision rapid classification of ridge waveguide modes was achieved. By learning the mapping relationship between the geometric structure of the ridge waveguide and the single-mode characteristics, the network can predict the single-mode properties of unknown ridge waveguide geometric parameters. The results show that 99% classification accuracy is achieved by the model with only 252 training data. The classification time of a single data point is reduced from 52 seconds of the traditional time-domain finite difference algorithm to 5 milliseconds, and the computational efficiency is improved by about 1 000 times. At the same time, the model has good generalization ability. This algorithm is expected to shorten the design cycle of optical waveguides, which has a certain supporting role in the design and preparation of large-scale photonic chips.