Beam Characterization For Few-mode Fibers Based On Deep Learning

The beam characterization including mode decomposition and beam quality determination is an important tool for the further research on fiber lasers.Deep learning technology,which has attracted much attention in recent years,is promising to provide a simple,effective,fast and accurate characterization scheme.This paper focues on how to utilize deep learning to characterize the beam emitting from few-mode fibers.We have proposed and verified the deep learning-based mode decomposition for few-mode fibers for the first time,to the best of our knowledge.Large amounts of samples are prepared for the convolutional neural network(CNN),helping it learn the relationship between the beam pattern and the mode coefficient.The experiment was carried out and the results have shown that the decomposing rate of our scheme can reach as high as 30 Hz.Besides,the imperfected patterns,including the noisy and saturated ones,can also be accurately decomposed by the CNN,which is more robust than traditional methods.The beam quality determination based on deep learning is also investigated for the first time.The trained CNN can predict the beam quality accurately with only single near-field beam pattern.Besides,the patterns with vertical blooming are researched thorugh simulation and the results exhibit high accuracy of CNN.Compared with the traditional schemes,the cost time of deep learning-based scheme is greatly reduced.Based on the achievements in mode decomposition and beam quality determination of few-mode fibers,we successfully extend the deep learning technique to the phase control of coherent beam combining systems.The phase error of the beam array could be estimated and compensated directly with the help of CNN,providing a new supplementary of phase control method without iterative process.