Tri‐Layer Laminated HZO/HfO ₂ /HZO Ferroelectric Diode for Linear Weight Updates in Hardware‐Aware Reservoir Computing

ABSTRACT This study proposes an Hf 0.5 Zr 0.5 O 2 (HZO)/HfO 2 /HZO tri‐layer laminated ferroelectric diode as a synaptic device platform and quantitatively demonstrates how device‐level domain‐switching dynamics translate into weight‐update linearity and, ultimately, system‐level classification performance. The tri‐layer laminated stack suppresses leakage path formation along vertical grain boundaries, reduces the off‐current, and enables a larger memory window and an enhanced tunneling electro resistance ratio. Incremental positive‐up‐negative‐down (PUND) measurements further reveal gradual, spatially distributed partial polarization switching over a wide voltage range in the laminated device, consistent with a more subdivided, multi‐domain switching process. Leveraging these characteristics, potentiation and depression behaviors were evaluated in a quasi‐stable regime after short‐term relaxation, confirming a pronounced improvement in weight‐update linearity for the laminated structure. Furthermore, we implemented a hardware‐aware reservoir computing framework that periodically snaps readout weights to experimentally measured potentiation and depression states under identical reservoir mapping. Using the laminated device dataset, the RC system achieves high accuracy on a 20‐class hand gesture recognition task with more uniform class‐wise performance. Overall, experiments and simulations show that laminated stack engineering enables gradual polarization switching, improves analog weight updates, and enhances neuromorphic performance, highlighting laminated ferroelectric diodes as a promising device‐algorithm co‐design platform for analog in‐memory computing and sequence‐based edge intelligence.