Deep learning-based classification of high intensity light patterns in photorefractive crystals

In this paper, we establish a new scheme for identification and classification of high intensity events generated by the propagation of light through a photorefractive SBN crystal. Among these events, which are the inevitable consequence of the development of modulation instability, are speckling and soliton-like patterns. The usual classifiers, developed on statistical measures, such as the significant intensity, often provide only a partial characterization of these events. Here, we try to overcome this deficiency by implementing the convolution neural network method to relate experimental data of light intensity distribution and corresponding numerical outputs with different high intensity regimes. The train and test sets are formed of experimentally obtained intensity profiles at the crystal output facet and corresponding numerical profiles. The accuracy of detection of speckles reaches maximum value of 100%, while the accuracy of solitons and caustic detection is above 97%. These performances are promising for the creation of neural network based routines for prediction of extreme events in wave media.