Self-rectifying memristor crossbar arrays enabling energy-efficient multiply–accumulate for on/off-chip neuromorphic computing

Self-rectifying resistive random-access memory (SRRAM) offers intrinsic diode-like behavior that suppresses reverse current flow, enabling selector-free operation and reducing power consumption in high-density crossbar arrays. Here, we present a silicon-doped zirconium oxide based SRRAM device exhibiting high selectivity (10⁶), robust endurance (> 2 × 10⁶ cycles), stable data retention (7200 s at 85 °C), and well-resolved multilevel conductance states. The device was implemented in a 32 × 32 crossbar array using a 1/3 biasing scheme, achieving accurate and energy-efficient multiply-and-accumulate (MAC) operations at the array level. To address device-level variability, uncertainty-aware learning was performed through propagation modeling, enabling reliable hardware-aware handwritten digit classification with 92% accuracy. Furthermore, the SRRAM array supported off-chip inference through device-characteristic-aware mapping of the first convolutional layer in a GAN-based image reconstruction model, demonstrating high-fidelity analog computation in vision tasks. These results establish Si-doped ZrO x SRRAM as a reliable and energy-efficient synaptic platform capable of supporting hardware-aware neuromorphic inference and off-chip inference for neuromorphic computing applications.