Pulse‐Engineered Synaptic Linearity and Non‐Volatile Memory in MoO ₃ /TiO ₂ Bilayer Memristors for Neuromorphic Image Recognition

ABSTRACT

Memristive devices integrating non‐volatile memory and artificial synaptic functionalities are promising candidates for neuromorphic in‐memory computing. However, stochastic and non‐linear evolution of conductive filaments often hinders stable and gradual synaptic weight updates. Here, we investigate pulse‐protocol optimization in an Au/MoO ₃ /TiO ₂ /FTO bilayer‐oxide memristor to achieve both digital memory operation and analog conductance modulation. The switching characteristics evolve from interface‐driven self‐rectifying behavior prior to electroforming, to filament‐governed non‐volatile bipolar resistive switching. The device exhibits good endurance (25,000 cycles) and retention (1000 s), fast switching, low variability, operational stability up to 60°C, and reliable performance over a testing duration of 70 weeks. Furthermore, the RESET process exhibits quantized conductance features, indicating atomic‐scale constricted conduction pathways across the switching layers. Importantly, the optimized combinational voltage pulse scheme enables gradual and controlled conductance evolution, resulting in highly linear and stable potentiation– depression characteristics. The pulse‐engineered synaptic linearity further enables a significant ∼20% enhancement in artificial neural network inference accuracy for handwritten digit recognition. This work highlights pulse‐protocol control combined with oxide‐bilayer architecture as an effective strategy for developing multifunctional memristors for neuromorphic computing applications.