Material-engineered self-compliant memristor enabling multibit synaptic learning and in-memory computing

Abstract

To improve the integration density of memristor-based crossbar arrays, we present a self-compliant RRAM device that eliminates the need for external current-limiting transistors. Additionally, an Al2O3 layer enhances switching uniformity by stabilizing the filament path, while an AlN layer acts as an oxygen barrier to improve retention. A thin SiO2 tunnel barrier is also introduced to modulate oxygen ion migration, which localized the formation of conductive filaments and significantly improved endurance by reducing cycle-to-cycle variation. This material-engineered architecture enables stable and repeatable resistive switching with low variability and high endurance. Neuromorphic characteristics were evaluated by applying pulse-based electrical stimuli to emulate biological synaptic behaviors such as long-term potentiation (LTP), long-term depression (LTD), and spike-timing-dependent plasticity (STDP), enabling analog synaptic weight updates. The multibit capabilities of the device were systematically investigated by modulating voltage amplitudes and applying incremental step pulse with verify algorithm (ISPVA) scheme, demonstrating reliable conductance tuning up to 8-bit resolution. Furthermore, the device was integrated into a memristor-based pattern recognition system for edge-computing-oriented neuromorphic inference, where externally controlled conductance states were used to emulate synaptic weights during digit classification. These results highlight the potential of the proposed memristor array as a scalable and energy-efficient platform for on-device learning and in-memory computing applications in edge environments.