Abstract: In order to solve the problem of detection of different orbital angular momentum (OAM) superimposed vortex beams, a pattern recognition technology based on machine learning (ML) is proposed, which provides a brand-new method for multi-OAM states detection. In order to study the recognition rate of multi-OAM beams using convolutional neural network (CNN) models under different wavelength, transmission distance and atmospheric turbulence conditions, the numerical simulation phase screens are generated by the power spectral inversion method based on the modified von Karman power spectrum model. Multi-step diffraction method is used to simulate numerically the propagation of OAM beams in the atmospheric turbulence, and the training and testing database are obtained under different atmospheric turbulence. Results indicate the accuracy of CNN-based OAM pattern recognition increases as wavelength increases, transmission distance decreases and turbulent intensity decreases. And the CNN trained with the database under strong turbulence has high accuracy for all kind of turbulence condition, and using mixed training database under different turbulence condition can enhance the accuracy. These results contribute to the demultiplexing systems of free space optical-OAM systems.

Key words: orbital angular momentum, atmospheric turbulence, spiral phase spectrum, convolutional neural network