Machine learning algorithms predict experimental output of chaotic lasers

Abstract

We apply an artificial neural network (ANN) of 20 hidden layers and backpropagation regression to the forecast of experimental time series from a KLM Ti:Sapphire laser and a Nd: Vanadate with modulation losses. In both cases the neural network is able to predict up to 10 steps ahead. In the Ti:Sa laser the prediction in pulse amplitude is accurate even when the pulse is an extreme event. In the Nd: Vanadate laser we forecast both pulse amplitude and pulse to pulse time separation. In both cases the prediction goes beyond the Lyapunov prediction horizon.