Hua Shuaibin, Zhang Le, Wang Liang, Zheng Ruhui, Gan Puli, Guo Xin
State Key Laboratory of Material Processing and Die & Mould Technology, School of Materials Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074, P. R. China.
Engineering Research Center for Functional Ceramics of Ministry of Education, School of Integrated Circuits, Huazhong University of Science and Technology, Wuhan 430074, P. R. China.
ACS Appl Mater Interfaces. 2025 Sep 24;17(38):53691-53703. doi: 10.1021/acsami.5c10190. Epub 2025 Sep 11.
Reservoir computing (RC) systems have a time signal processing architecture with the advantages of high efficiency and low training cost. Oxide-based memristors present a promising solution for the development of high-performance, scalable, memristive reservoir computing systems, benefiting from their inherent dynamic nonlinearity and substantial commercial potential. Compared to conventional thin-film deposition techniques, the anodization technique demonstrates advantages in cost-effectiveness, processing speed, and operational simplicity in preparing oxide films for memristors. However, anodized memristors usually have limited device structures, and their nonvolatile characteristics are incompatible with the RC systems. In this study, TiN/NbO/Pt memristors with low cycle temporal variation (<5%) and high yield are prepared via the anodization technique at 40 s. The resistive mechanism of memristors has been systematically investigated, and the devices have been modeled accordingly. Then, compression of MNIST images in both horizontal and vertical dimensions is achieved through memristors. Compared to the original data, the training time is reduced by 86.8% while ensuring the classification accuracy (97.25%). The memristor-based reservoir computing network exhibits good prediction of Hénon map sequences at the simulation and hardware level with an average power consumption as low as 1.97 μW for a single pulse.
储层计算（RC）系统具有时间信号处理架构，具有高效和低训练成本的优点。基于氧化物的忆阻器为高性能、可扩展的忆阻性储层计算系统的发展提供了一个有前景的解决方案，这得益于其固有的动态非线性和巨大的商业潜力。与传统的薄膜沉积技术相比，阳极氧化技术在制备用于忆阻器的氧化膜时，在成本效益、处理速度和操作简便性方面具有优势。然而，阳极氧化忆阻器通常具有有限的器件结构，并且它们的非易失性特性与RC系统不兼容。在本研究中，通过40秒的阳极氧化技术制备了具有低循环时间变化（<5%）和高成品率的TiN/NbO/Pt忆阻器。系统地研究了忆阻器的电阻机制，并据此对器件进行了建模。然后，通过忆阻器实现了MNIST图像在水平和垂直维度上的压缩。与原始数据相比，在确保分类准确率（97.25%）的同时，训练时间减少了86.8%。基于忆阻器的储层计算网络在模拟和硬件层面上对Hénon映射序列表现出良好的预测能力，单个脉冲的平均功耗低至1.97μW。
