Performance Maximization of In-Memory Reinforcement Learning with Variability-Controlled Hf1-xZrxO2 Ferroelectric Tunnel Junctions

We develop strategies to maximize the performance and reliability of in-memory reinforcement learning (RL) with Hf
1-x
Zr
x
O
2
(HZO) ferroelectric tunnel junction (FTJ) by structural and material engineering. Simulation results of RL suggest that small cycle-to-cycle variability is desirable for effective policy finding while large operation voltage margin (V
m
), which is the difference between breakdown voltage (V
BD
) and set voltage (V
set
), facilitates V
set
tuning for performance maximization. HZO thinning is effective to reduce the variability as a result of reduced effect of traps on the current, while V
m
can be increased with decreasing Zr concentration. We also found that device size scaling enables low-power, high-density, and high-reliability RL with keeping the learning performance.