Leveraging Tunability\nof Localized-Interfacial Memristors\nfor Efficient Handling of Complex Neural Networks

A scalable (<130 nm) resistive switching memristor\nthat features\nboth filamentary and interfacial switching aimed at neuromorphic computing\nis developed in this study. The typically perceived noise or volatility\nwas effectively harnessed as a controlled mechanism for interfacial\nswitching. The multilayer structure for the proposed memristor enhances\nswitching stability by curbing ionic overmigration and mitigating\nleakage paths. Furthermore, the memristors showcased their reliability\nby demonstrating more than 15 M cycles in the filamentary mode and\n1 M pulses in the interfacial mode. Additionally, retention tests\nat 85 °C for 10<sup>4</sup> s confirmed the stability across\ndifferent states, affirming its reliability as a nonvolatile CMOS-compatible\nelement. While many studies validate performance solely on the MNIST\ndata set, this work also evaluates more complex data sets, demonstrating\nthe robustness of the demonstrated memristor in supervised learning.\nSpecifically, supervised learning simulations on MNIST and fashion\nMNIST data sets indicated a high learning rate with <4% deviations\nfrom numerical training, while offline inference trained on CIFAR-10\nand CIFAR-100 data sets revealed <2.5% and <7% deviations caused\nby programing error accumulation, even with increased memristor counts\nfor these highly complex data sets. Unsupervised learning via spike-timing-dependent\nplasticity further highlights the potential of the developed memristor\nin bridging artificial and biological paradigms, offering a significant\nadvance toward efficient and biologically inspired computing architectures.