Back-End-Of-Line Integration of 2-Terminal Ferroelectric Hafnia Synaptic Weights on Commercial CMOS Process

Abstract

Functionalising the back-end-of-line (BEOL) of CMOS with emerging materials is a key route to neural network accelerators and neuromorphic systems. In this work, we demonstrate for the first time the integration of 2-terminal ferroelectric synaptic weights in the BEOL of as-delivered commercial 180 nm CMOS dies. A novel material stack is proposed, in which a WOx/[HfO2-ZrO2] bilayer is sandwiched between TiN and Mo electrodes. For devices annealed at 450 °C, programming below 1.8 V is achieved with a cycle-to-cycle variation below 3 %. Synaptic weights annealed at 500 °C exhibit more gradual weight updates, while the tested underlying NMOS transistors remain functional after the complete integration process. The trade-off between thermal budget and on-device training performance is illustrated using hardware-aware neural network simulations.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.