Optimization of photonic reservoir computing using short internal cavity laser

Abstract

With the advent of the high-speed information age and the explosive growth of the information, higher requirements have been placed on the information processing speed. In recent years, the delay-based reservoir computing (RC) systems have been extensively investigated. Meanwhile, the information processing rate is improved mainly around the replacement of nonlinear nodes in the system. Nevertheless, as the most commonly used distributed feedback semiconductor (DFB) laser, many researchers only use ordinary commercial DFB products for research, and they have not noticed the improvement of RC performance caused by changes in internal parameters of laser. With the development of photonic integration technology, the processing technology of DFB turns more mature, so that the size of DFB can be fabricated in a range of 100 μm–1 mm when it still generates laser, and the photon lifetime of the laser will also change. The shorter photon lifetime in the laser leads to a faster dynamic response, which has the potential to process the information at a higher rate in the RC system. According to the laser rate equation (Lang-Kobayashi), changing the internal cavity length will affect the feedback strength, injection strength and other parameters required for the laser to enter into each dynamic state, which in turn affects the parameter space required for the RC system to exhibit high performance. According to this, we study the relationship between the internal cavity length (120 μm–900 μm) and the information processing rate of the RC system. In addition, the influences of different internal cavity lengths on the parameter space of the RC system are analyzed. The results show that when the internal cavity length is in a range from 120 μm to 171 μm, the system can achieve 20-Gbps low-error information processing. It is worth noting that when the internal cavity length decreases from 600 μm to 128 μm, the parameter space with better prediction performance of the RC system is greatly improved. When performing the Santa-Fe chaotic time series prediction task, the normalized mean square error (NMSE) is less than 0.01, and the parameter range of the injection strength is increased by about 22%. The range of parameter with NMSE no more than 0.1 is improved by nearly 40% for the 10^th order nonlinear auto-regressive moving average (NARMA-10) task. When the number of virtual nodes is 50, the system can achieve a high-precision prediction for the above two tasks. This is of great significance for the practical development of the system.