神经形态工程学
油藏计算
光电子学
材料科学
计算机科学
编码(内存)
MNIST数据库
晶体管
特征(语言学)
联轴节(管道)
电子工程
核(代数)
等离子体子
混合动力系统
频道(广播)
编码(集合论)
光电导性
硅
电介质
光子学
电极
数码产品
国家(计算机科学)
调制(音乐)
光伏系统
逻辑门
人工神经网络
作者
Dongdong Sun,Shaofan Zhao,Liuliu Jiang,Gongpeng Lan,Xinyang Deng,Wuqing Fang,Shaochuan Chen,Yuanyuan Shi
标识
DOI:10.1088/1361-6463/ae39cb
摘要
Abstract Physical reservoir computing has emerged as an energy-efficient neuromorphic paradigm for processing complex temporal data. However, conventional single-modality reservoirs often suffer from limited internal state richness, constraining their capability to handle high-dimensional tasks. Here, we report a multimodal optoelectronic reservoir computing system based on ZnO thin-film transistors. By exploiting the synergistic interaction between oxygen ion kinetics in the HfO 2 gate dielectric and the persistent photoconductivity effect in the ZnO channel, we identify a robust asymmetric optoelectronic coupling mechanism. Specifically, the electrical gate bias nonlinearly modulates the optical relaxation dynamics, effectively acting as a tunable kernel to expand the reservoir’s feature mapping space. Leveraging this property, we implement a hybrid optoelectronic encoding strategy that achieves a peak recognition accuracy of 90.31% on the MNIST handwritten digit dataset, significantly outperforming single-modality benchmarks. Our analysis reveals that the asymmetric cross-coupling breaks the monotonicity of the device response, thereby enhancing state separability. This work provides a fundamental physical framework for leveraging single-device multimodal dynamics to develop high-performance neuromorphic hardware.
科研通智能强力驱动
Strongly Powered by AbleSci AI