Improved State Stability Of Hfo2 Ferroelectric Tunnel Junction By Template-Induced Crystallization And Remote Scavenging For Efficient In-Memory Reinforcement Learning

We investigated the effects of read current instabilities originated from depolarization field and charge trapping in HfO2 ferroelectric tunnel junctions (FTJs) on the performance of in-memory reinforcement learning. We utilized, for the first time, remote scavenging to control interfacial layer thickness, combined with template-induced crystallization to stabilize the ferroelectric phase. These are found to improve both short-term and long-term stability of memory state. Pole-cart simulation results reveal that these improvements significantly contribute to the efficiency and stability of reinforcement learning with the FTJ cross-point array.