Impact of the Barrier Layer on the High Thermal and Mechanical Stability of a Flexible Resistive Memory in a Neural Network Application

Abstract

An artificial synaptic device with continuous conductance variation is essential for the hardware implementation of bioinspired neuromorphic systems. With increasing technological advancement, wearable flexible technology is gaining enormous importance in the research community. High temperature is one of the key issues in flexible technology from the fabrication and applicability aspects. In this work, we have demonstrated the performance of a complementary metal− oxide−semiconductor (CMOS)-compatible high-k material (HfO 2 ) based flexible heterogeneous stacked resistive switching device in an artificial neural network. The device exhibited an excellent memory window (I ON /I OFF ) of around 1.2 × 10 4 with an ultralow variation (σ HRS ∼ 3.95 × 10 −10 S) at 500 μA current compliance. The device shows excellent mechanical and electrical stability for more than 500 DC cycles and data retention capability at 120 °C for more than 10 4 s without any degradation. The device can be used for multilevel cell (MLC) operation in six distinct states and can be useful for the implementation of 2-bit data storage applications. The conduction mechanism in the device was dominated by Schottky emission at the lower field region and hopping conduction at the higher field region of the high-resistance state (HRS), whereas ohmic conduction was satisfied at the lowresistance state (LRS). We have trained the device in the neural network with 96.07% accuracy as the baseline and observed the effect of conductance variation and high-temperature operation. We trained the device at a high temperature with a 95.68% off-chip training accuracy and observed the accuracy profile throughout the time. The device also possessed an excellent mechanical stability under a long-term bending stress (r = 8 mm) of over 10 4 s with an intact memory window. The neural accuracy was measured every 30 min with a maximum of 96.01% to observe the effect of long-term mechanical stress on the off-chip learning process.

Flexible Memristor-Based Nanoelectronic Devices for Wearable Applications: A Review

“…In 2022, they further constructed a flexible memristor with the Ni/Ti/ HfO 2 /Ni/Ti/SiO 2 structure on polyimide. 83 The device with two Ti buffer layers has superior mechanical stability at a bending radius of 8 mm, which is more favorable for neuromorphic computation, as shown in Figure 6c,d. The buffer layers influence the formation and rupture of the conducting filaments and improve the stability of the device.…”

“…More importantly, the device has excellent flexibility at a curvature radius of 5 mm. In 2022, they further constructed a flexible memristor with the Ni/Ti/HfO 2 /Ni/Ti/SiO 2 structure on polyimide . The device with two Ti buffer layers has superior mechanical stability at a bending radius of 8 mm, which is more favorable for neuromorphic computation, as shown in Figure c,d.…”

“…(d) Comparative analysis of the neural efficiency of flexible devices under bending. Reproduced with permission from ref . Copyright 2022 American Chemical Society.…”

“…It has been found that embedding a thin layer between the electrode and functional layer can effectively improve the performance of memristors. − In 2020, Kumar et al prepared a flexible memristor with the TiN/Al 2 O 3 /ZnO/Al 2 O 3 /TiN structure . The addition of two layers of Al 2 O 3 with thickness of about 3 nm gives the device more uniform and stable bipolar RS characteristics, and the device has a resistance ratio of more than 2 orders of magnitude.…”

Controllable coexistence of threshold and non-volatile crosspoint memory for highly linear synaptic device applications

“…The eNVM-based in-memory neuromorphic computing architectures based on the weight update rule through the synapses similar to the processing of information in biological neurons provides a new direction to perform power-efficient artificial neural networks (ANNs) with much higher correspondence and less data processing than conventional systems [10,11]. Amongst the various eNVMs, the resistive random access memory (RRAM) has the superiority to satisfy the demands for the next-generation hardware-implemented neuromorphic computing and Internet of Things (IoT) because of its non-destructive readout, highdensity integration, fast operation speed, low power consumption, excellent scalability and good compatibility with CMOS technology [12][13][14]. The excellent scalability of RRAM facilitates the fabrication of large cross-point arrays, which are essential for highly efficient communication amongst complex neural networks [15].…”