Linear and Programmable Long‐Term Plasticity in PECVD Amorphous SiC Memristors

ABSTRACT Amorphous silicon carbide (SiC) is a promising dielectric platform for memristive artificial synapses due to its high stability and back‐end‐of‐line compatibility for CMOS integration. Here, we demonstrate stoichiometric PECVD‐grown amorphous SiC memristors with Cu active electrodes, in which synaptic functionality is engineered through control of material composition. By tuning the SiC stoichiometry from Si‐rich to near‐stoichiometric, the bonding configuration is modified to suppress excessive Cu ion migration, enabling gradual and stable conductance evolution. The resulting devices exhibit highly linear and programmable long‐term potentiation and depression, with a low nonlinearity of 0.08, uniform conductance updates, and minimal cycle‐to‐cycle variation. The synaptic behavior is further shown to be controllable through pulse programming, enabling tunable multilevel conductance states. Using experimentally extracted device characteristics, we implement both artificial neural networks and convolutional neural networks to evaluate system‐level performance. The memristor‐based networks achieve classification accuracies exceeding 95% on MNIST and 90% on CIFAR‐10, while maintaining robust performance under conductance discretisation and cycle‐to‐cycle variation. These results demonstrate a materials‐driven approach to achieving linear and controllable synaptic behavior, establishing PECVD amorphous SiC as a strong candidate for high‐precision and scalable neuromorphic computing systems.