Gradient Nitrogen-Doped Memristor for Computing-in-Memory.

Oxide-based memristors hold promise for computing-in-memory (CIM) architectures due to their compatibility with complementary metal-oxide-semiconductor processes. However, the intrinsically nonuniform growth rates and stochastic formation pathways of conductive filaments (CFs) degrade conductance modulation linearity and cycle-to-cycle uniformity, thereby hindering their application in reliable CIM systems. Here, we propose a simple approach to simultaneously improve the linearity and uniformity of HfO memristors through gradient nitrogen-doping (GND). This doping strategy establishes a stair-like energy barrier profile for oxygen ion migration, effectively regulating the growth rate of CFs and thereby enhancing linearity. Concurrently, nitrogen dopants confine CF growth along the same pathways during repeated cycles, thereby achieving ultrasmall switching voltage variation (3.1%) and resistance variation (5.7%). Furthermore, by successfully executing stateful Boolean logic and implementing long short-term memory network cores in GND memristor arrays for high-accuracy temporal sequence prediction tasks, the GND memristor demonstrates its versatility in both digital and analog CIM systems.