Achieving Robust Multistate Memristors Through Sparsely Embedding Quantum Dots for Neuromorphic Applications

Artificial intelligence computing with contemporary hardware demands substantial energy consumption. Memristors, with their inherent in-memory computing capabilities, offer a promising solution. However, their application is hindered by challenges such as non-linearity and limited cycling endurance for multi-state performance, primarily due to the stochastic evolution of conductive filaments. In this work, lead sulfide quantum dots are embedded in SiO2-based memristors to avoid abrupt conductance jumps and increase the cycling numbers of long-term potentiation and depression (LTP/LTD) under continuous pulse stimuli. Compared with pure SiO2 and densely embedded quantum dot memristors, those with sparsely distributed quantum dots achieved gradual multistate switching and hundreds of LTP/LTD cycles. Transmission Electron Microscopy and Multiphysics studies indicate that these sparse quantum dots induce unevenly distributed electrical fields, in turn creating favorable locations for the evolution of conductive filaments, resulting in significant performance enhancements. We also further improved the linearity with a programmed pulse scheme. These optimized synaptic characteristics were then applied to neural networks for image recognition tasks, using the MNIST and Fashion MNIST datasets, achieving inference accuracies comparable to those obtained with conventional computing methods. In essence, this work provides a transformative strategy to engineer memristors for achieving robust multistate performance towards neuromorphic computing applications.