Prediction of optical chaos based on a liquid neural network

This paper investigates the prediction of chaotic time series in semiconductor lasers using a neural network based on liquid time-constants (LTC), which incorporates a neural circuit policies (NCP) and a long short-term memory (LSTM) module to capture both the long-time and short-time characteristics. Compared to traditional recurrent neural networks, this method features variable time constants that influence the network's decision-making speed while mimicking biological neural systems. Simulation results show that in the prediction task, only a small number of neurons are required to achieve good prediction performance. Moreover, continuous prediction can be realized through the cyclic operation of training and testing, with the longest prediction duration reaching 9 ns.