WTF-former: A model for predicting optical chaos in laser system

Abstract

Optical chaotic phenomena in semiconductor lasers have a wide range of applications in random bit generation and secure communications, and although a lot of effort has been expended to study these chaotic behaviors through numerical simulations, it is still challenging to accurately predict chaotic dynamics using a limited number of observations. Existing solutions often start from time-domain features only and use machine learning and neural network means to make predictions, with insufficient ability of the model to learn chaotic behaviors and low prediction accuracy. Here, we propose a novel neural network model for continuous prediction of optical chaotic phenomena, the WTF-former, which has a strong ability of chaotic feature learning and can effectively capture the behavior of optical chaos. It designs an iTransformer neural network based on Wavelet Convolution (WTConv) and Frequency Enhanced Channel Attention Mechanism (FECAM), which can capture the feature information in both time and frequency domains for better prediction of optical chaotic phenomena. The simulation results show that our proposed model is capable of predicting chaotic time series up to 6 ns under different laser conditions, and the Mean Absolute Error (MAE), Root Mean Squard Error (RMSE), and R-Square ($R^2$) comparative metrics are better than those of existing models. Finally, we conducted ablation experiments to verify the effectiveness of our designed WTConv, FECAM module, and iTransformer as the backbone network. The results show that our designed WTF-former network is capable of accurate and continuous prediction of semiconductor laser optical chaos phenomena over a long period of time.