Two-step hybrid retrieval algorithm for characterizing ultrashort optical pulse from its SHG-FROG trace

Abstract

In this paper, we propose a two-step hybrid retrieval algorithm for characterizing ultrashort optical pulse from its measured SHG-FROG trace. In the first step, we use a modified ResNet50 deep neural network and a very small training datasets containing only 4000 FROG-trace samples to quickly obtain the rough characteristics of the pulse to be retrieved. Then, we use the results obtained in the first step as the initial values of the iterative optimization in the second step, which can further obtain more accurate characteristics of ultrashort optical pulse. Results demonstrate that the two-step hybrid retrieval algorithm can achieve very high accuracy in characterizing ultrashort optical pulses. And the correlation coefficient between the reconstructed FROG trace based on the retrieved ultrashort optical pulse and the original measured FROG trace can approximately reach 1.