Sodium-enhanced perovskite reservoir for photonic in-sensor computing

Abstract

Physical reservoir computing (RC) mimics the brain's capacity for temporal processing by mapping inputs into high-dimensional feature spaces. This biologically inspired approach offers advantages in terms of training efficiency and real-time performance. To enhance physical RC's accuracy, nonlinear and dynamic responses are essential to distinguish complex time-series input data. Here, we present a sodium-doped perovskite memristor-based RC system that capitalizes on the synergistic effects of photovoltaic and photogating to induce nonlinear and high-dimensional dynamics. By precisely controlling the sodium doping concentration, we achieve a wide range of distinct conductance states (16 levels), enabling the system to effectively process diverse temporal patterns. We demonstrate the system’s capabilities across a range of computational tasks, achieving a 92.11 % accuracy in image recognition and a low normalized root-mean-square error (NRMSE) of 0.056 in temporal Hénon map prediction. Our findings demonstrate the potential for future development of high-performance memristor-based RC systems, particularly those capable of handling complex temporal tasks.