Scaled, Ferroelectric Memristive Synapse for Back‐End‐of‐Line Integration with Neuromorphic Hardware

Abstract

Ohmic, memristive synaptic weights are fabricated with a back‐end‐of‐line compatible process, based on a 3.5 nm HfZrO4 thin film crystallized in the ferroelectric phase at only 400 °C. The current density is increased by three orders of magnitude compared to the state‐of‐the‐art. The use of a metallic oxide interlayer, WOx, allows excellent retention (only 6% decay after 106 s) and endurance (1010 full switching cycles). The On/Off of 7 and the small device‐to‐device variability (<5%) make them promising candidates for neural networks inference. The synaptic functionality for online learning is also demonstrated: using pulses of increasing (resp. constant) amplitude and constant (resp. increasing) duration, emulating spike‐timing (resp. spike‐rate) dependent plasticity. Writing with 20 ns pulses only dissipate femtojoules. The cycle‐to‐cycle variation is below 2%. The training accuracy (MNIST) of a neural network is estimated to reach 92% after 36 epochs. Temperature‐dependent experiments reveal the presence of allowed states for charge carriers within the bandgap of hafnium zirconate. Upon polarization switching, the screening of the polarization by mobile charges (that can be associated with oxygen vacancies and/or ions) within the ferroelectric layer modifies the energy profile of the conduction band and the bulk transport properties.

From Ferroelectric Material Optimization to Neuromorphic Devices

“…conduction [164] or modified Schottky emission, [165] which are also nonlinear conduction mechanisms, making them poor candidates for the analog implementation of the "multiply" operation. Berdan et al exploited the fact that the nonlinearity factor was constant for various conductance levels in ferroelectric Si : HfO /SiO 2 2 tunnel junctions: as represented in Figure 6a, the nonlinearity of the current-voltage relation was circumvented by logarithmic line drivers.…”

“…Increasing the thickness of the ferroelectric layer leads to an increase in the relative contribution of the thermionic emission. [ 163 ] In hafnia ferroelectrics, the presence of shallow traps is associated to Poole–Frenkel conduction [ 164 ] or modified Schottky emission, [ 165 ] which are also nonlinear conduction mechanisms, making them poor candidates for the analog implementation of the “multiply” operation. Berdan et al.…”

“…Reproduced with permission. [ 165 ] Copyright 2022, John Wiley and Sons. c) Drain (D)–Source (S) conductance measured after programming a HZO/HfO 2 /HZO/IWO FeFET device using pulses of constant duration and linearly increasing amplitude, as represented in the inset.…”

An efficient deep neural network accelerator using controlled ferroelectric domain dynamics

“…This can cause significant acceleration in training and improvement in training energy. In perovskite-oxide-based FTJs [28], Hf 0.5 Zr 0.5 O 2 (HZO)-based FTJs [29], FeFETs [13], Fe-thin film transistors (FeTFTs) and Fe-finFETs [15,16], efficient DNN training has been demonstrated. The FeFETs, Fe-finFETs and FeTFTs (one-transistor memory) have a larger dynamic range (on/off ratio exceeding 10 6 or similar) to operate on, which improves the training accuracy.…”

On-Chip Integrated Atomically Thin 2D Material Heater as a Training Accelerator for an Electrochemical Random-Access Memory Synapse for Neuromorphic Computing Application

“…Advances in memory technology have yielded numerous types of two memory devices for neuromorphic hardware. − However, due to excessive writing nonlinearity and noise, as well as high switching voltages in two-terminal memristors, efficient analog nonvolatile memory components (artificial synapses) remain elusive. Three-terminal electrochemical random-access memory (ECRAM) has better synaptic characteristics and lower energy consumption than traditional memory devices. − In ECRAM, the synaptic weights are stored in the form of analog conductance change in channel layer.…”