Complementary Self-Selective Memristor for Intrinsic Differential Behavior

Abstract

Memristor-based in-memory computing relies on differential pairs for signed weights and suffers from sneak paths, large area, and device mismatch. Here, we present a novel memristor device with intrinsic differential characteristics enabling bidirectional weight tuning, offering a promising pathway toward high-efficiency in-memory computing with great potential for in-situ learning. The device exhibits a self-rectification ratio and nonlinearity of $10^{{5}}$ under both polarities, with an endurance exceeding $10^{{6}}$ cycles. Under pulse programming, bidirectional potentiation and depression are realized, enabling intrinsic positive and negative conductance modulation within a single device. By tuning the bidirectional conductance asymmetry via pulses, single-device-level differential computation is achieved, suppressing conductance drift and circumventing the inherent non-uniformity of dual-device architectures. We experimentally achieve low-error differential weight updates on a ${32}\times {32}$ crossbar array. The device’s native bipolar conductance supports differential readout, potentially removing the need for independent subtraction hardware and providing a novel optimization route for in-situ learning CIM.