Sequential Process Optimization of HfO₂/Al₂O₃ Bilayer RRAM for Enhanced Synaptic Performance and On-Chip MNIST Learning

HfO₂/Al₂O₃ bilayer resistive random-access memory (RRAM) exhibits gradual resistive switching behavior, which is advantageous for achieving linear and symmetric conductance modulation required for reliable synaptic operation in neuromorphic computing. However, large device variability and poor endurance remain critical challenges that must be addressed for practical synaptic applications of RRAM. In this work, the performance of HfO₂/Al₂O₃ bilayer RRAM was systematically improved through the sequential optimization of fabrication conditions. By combining postdeposition annealing (PDA) of the switching layer, optimization of the titanium (Ti) buffer layer thickness, and ultrathin molybdenum (Mo) deposition at the HfO₂/Al₂O₃ interface, the optimized devices exhibited over 70% reduction in cycle-to-cycle (C2C) operational variability, excellent DC endurance (>10³ cycles), a high on/off ratio (average 96.8), and robust retention (>4000 s at 85 °C). Cross-sectional transmission electron microscopy and energy-dispersive X-ray spectroscopy (EDS) analyses revealed that Mo nanoislands formed by the ultrathin Mo layer play a key role in suppressing the stochastic formation of conductive filaments. In addition, fitting of DC I–V characteristics indicated that direct tunneling, Fowler–Nordheim tunneling, and Poole–Frenkel emission are the dominant conduction mechanisms governing device operation. Finally, leveraging the obtained long-term potentiation and long-term depression characteristics, on-chip learning-based pattern recognition was evaluated using the Modified National Institute of Standards and Technology (MNIST) data set, achieving a maximum classification accuracy of 81.56%. These results demonstrate the strong potential of performance-optimized HfO₂/Al₂O₃ bilayer RRAM as a synaptic device for neuromorphic computing applications.