Complete mode spectrum characterization for 1924 complex structured laser fields based on deep learning

Laser transverse mode analysis is of significance for laser applications. In beam quality analysis, the M² factor measurement is commonly used. However, it only reflects the transverse mode orders and can hardly represent the detailed information of laser field. In contrast, mode spectrum analysis provides a more comprehensive solution for laser mode analysis. At present, some research has utilized deep learning techniques to conduct mode spectrum analysis on vortex beams and Hermite—Gaussian (HG) beams. Nevertheless, there is a lack of research in the area of the recognition performance of deep learning when dealing with a large number of structured light beams. Here, we propose a novel multi—convolutional neural network integration method for analyzing the mode spectra of thousands of complex structured lights widely obtained from laser cavities. Based on the coherent superposition of HG eigenmodes, more than 1900 unique complex structured laser patterns are generated and divided into seven groups for parallel recognition by seven convolutional neural networks, effectively addressing the challenge of distinguishing highly similar mode patterns. The experimental results show that our multi—network scheme achieves high—precision mode spectrum analysis when analyzing 1924 simulated mode patterns with various distortions, with an accuracy rate exceeding 93%. It also demonstrates precise recognition ability for experimentally generated mode patterns. On the GPU RTX 4090D, the analysis time for each mode is approximately 11 ms. This method can obtain a complete analysis of the mode components of structured laser beams through only one intensity acquisition, providing a fast, simple, and cost—effective real—time mode spectrum analysis method for fields such as high—capacity optical communication, beam shaping, and beam quality evaluation.