Applications of machine learning to fiber laser optimization

Abstract

Neural network as a mathematical model has been applied in various fields to solve regression problems, ranging from economic prediction to inverse design of photonics systems. Powered by the multilayer interconnected structure and nonlinear activation functions, neural network can well model complex existent systems. In this work, we demonstrate that a feed-forward neural network (FNN) can be used to model the complex fiber laser cavity with nonlinear, dispersive and dissipative components, which are mutually dependent. Both gain and spectral filter bandwidth parameters are used as input of the neural network model to test the prediction capability of the trained models. Laser cavity parameters and laser outputs are correlated through the trained models. The results show that laser output spectrum and temporal pulse profiles can be predicted accurately with a normalized root mean square error (NRMSE) below 0.04. The prediction process can be completed within as short as 5 ms time frame. Additionally, we investigate the dependence of feed-forward neural network configuration, including layer amounts and neuron amounts, on prediction performance. We also showcase one example of ultrafast laser cavity inverse design from laser output to cavity parameters, enabling the architecting of on-demand intelligent lasers. The methodology here can be extended to various cavity structures and even different types of lasers advancing the field of real-world fast intelligent fiber laser cavity designing.