Authors: Zelin Ma, Huasen Yi, Ziping Zheng, Zhanyi Chen, Weicheng Liu, Yibing Chen, Bojun Cheng, Chang Cai, Shusheng Pan, Jun Ge
Published: 2025-05-14
Source: Full article
AbstractPhysical reservoir computing (PRC) holds great promise for low‐latency, energy‐efficient information processing, yet current implementations often suffer from limited flexibility, adaptability, and environmental stability. Here, a PRC system based on pulse‐width modulation (PWM)‐encoded resistor‐capacitor (R–C) circuits is introduced, achieving exceptional versatility and robustness. By leveraging customizable nonlinearities and dynamic timescales, this system achieves state‐of‐the‐art performance across diverse tasks, including chaotic time‐series forecasting (NRMSE = 0.015 for Mackey‐Glass) and complex multiscale tasks (94% accuracy for multiclass heartbeat classification). Notably, the design reduces relative errors by 98.4% across different device batches and under temperature variations compared to memristor‐based reservoirs. These features position the approach as a scalable, adaptive, and energy‐efficient solution for edge computing in dynamic environments, paving the way for robust and practical analog computing systems.