BEOL Large-Scale Integration and Precise Programming of HfOx/AlOy Superlattice- Like Multilevel Memristors

The memristor is a promising candidate for multilevel memory and neuromorphic computing. This study successfully integrated HfOx/AlOy superlattice-like (SLL) memristors in the 0.18/
$0.5~\mu $
m CMOS back-end-of-line (BEOL) process and statistically analyzed their performance. We identified that BEOL annealing (
$400~^{\circ }$
C/30 min) can cause oxygen diffusion and short-circuiting, but reducing Ti electrode thickness mitigates this issue. Furthermore, low post-forming resistance is discussed and solved by one transistor one memristor (1T1M) structure. Finally, we fabricate a 1T1M crossbar array with a size of 1 Kb in the
$0.18~\mu $
m CMOS process. We demonstrated the multilevel characteristics and endurance of 1T1M. Furthermore, the stable and precise multilevel conductance programming of memristors is achieved by a specific writing method and we simulated the effect of conductance of failure bits on classification accuracy in a four-layer MLP network on the MNIST dataset.