Time-delayed reservoir computing based on spin-VCSEL: interplay between pump ellipticity and performance

Reservoir computing, a simplified recurrent neural network, can be implemented using a nonlinear system with delay feedback, known as time-delayed reservoir computing. In this paper, we explore two time-delayed reservoir computing schemes based on the fast dynamics of two polarization channels of a spin-VCSEL and investigate their prediction performance for the Mackey–Glass task. Our main focus is on understanding the impact of pump ellipticity on the prediction performance of the two reservoir computing systems, namely, RC X and RC Y . Through numerical simulation, we find that when the pump ellipticity ( P ) is less than 0.73, the prediction performance of RC Y outperforms RC X . However, beyond this threshold, the performance advantage shifts towards RC X . These findings shed light on the importance of considering pump ellipticity when designing and optimizing reservoir computing systems. Furthermore, we also investigate the influence of the ratio between the delay time and input period on the memory capacity of these systems. Interestingly, we observe that using a delay time of 2.8 times the input cycle enables better prediction performance and memory capacity. This choice not only provides an optimal trade-off between memory capacity and computing speed but also avoids the computational slowdown caused by excessively long delay times. In general, our study emphasizes the flexibility and tunability of the spin-VCSEL-based reservoir computing system. By easily adjusting the ellipticity and delay-time parameters, we can optimize the memory properties, resulting in significantly improved prediction performance. Our findings offer valuable insights for enhancing the performance of reservoir computing systems based on the ultrafast dynamics of spin-VCSELs.