Zeng Xinlong, Xu Shule, Su Xiangwei, Zhang Cheng, Zhang Tianjiao, Wu Hongzhao, Wu Yan, Chen Yiqi, Xu Yang, Yu Bin, Liu Yang, Guo Yunfan, Guo Xiang, Xu Wei, Zhao Yuda
College of Integrated Circuits, ZJU-Hangzhou Global Scientific and Technological Innovation Centre, Zhejiang IC Innovation Platform, Zhejiang University, Hangzhou, 310027, China.
School of Computer Science and Technology, Zhejiang University of Science and Technology, Hangzhou, 310023, China.
Small. 2025 Aug 15:e05217. doi: 10.1002/smll.202505217.
Reservoir computing (RC) excels in temporal signal processing, driving advances in efficient reservoir hardware. However, dynamic target recognition faces challenges due to mismatches between event time scales and temporal properties of the optoelectronic RC system. In this work, a bridge is built between the event chronological information and the temporal dynamic of optoelectronic physical nodes in RC. The optoelectronic physical nodes are fabricated based on MoS phototransistors with varied fading memory timescale (τ) as the in-sensor RC hardware. Then the matching of τ with the time interval (Δt) of the input stimulus is explored by evaluating the linear separability (R) of reservoir states. When Δt/τ is within the range of 10-20%, the 32 output reservoir states from a 5-bit optical input display excellent linear separability with the R of 0.988 ± 0.006, contributing to the high accuracy rate of > 85.2% in recognizing eight sets of gestures. In comparison, when Δt/τ is out of the range of 10-20%, the R decreases and the recognition rate is below 77.6%. This study systematically quantifies the critical relationship between temporal scaling parameters and the time interval of optical input, providing a method to design the temporally adaptive optoelectronic physical nodes for high-efficiency in-sensor RC systems.
储层计算（RC）在时间信号处理方面表现出色，推动了高效储层硬件的发展。然而，由于事件时间尺度与光电RC系统的时间特性不匹配，动态目标识别面临挑战。在这项工作中，在RC中事件的时间顺序信息与光电物理节点的时间动态之间架起了一座桥梁。基于具有不同衰落记忆时间尺度（τ）的MoS光电晶体管制造光电物理节点，作为传感器内的RC硬件。然后，通过评估储层状态的线性可分性（R）来探索τ与输入刺激的时间间隔（Δt）的匹配。当Δt/τ在10%-20%范围内时，来自5位光输入的32个输出储层状态表现出优异的线性可分性，R为0.988±0.006，有助于在识别八组手势时获得>85.2%的高精度率。相比之下，当Δt/τ超出10%-20%范围时，R降低，识别率低于77.6%。本研究系统地量化了时间尺度参数与光输入时间间隔之间的关键关系，为设计用于高效传感器内RC系统的时间自适应光电物理节点提供了一种方法。
