Silicon based Bi_0.9La_0.1FeO₃ ferroelectric tunnel junction memristor for convolutional neural network application

Area of Science:

• Materials Science
• Solid State Physics
• Computer Engineering

Background:

• The von Neumann architecture faces performance bottlenecks due to data transfer limitations.
• Traditional conductive filament memristors exhibit unstable performance for advanced computing applications.
• Ferroelectric tunnel junctions (FTJs) offer a promising alternative for nonvolatile memory and neuromorphic computing.

Purpose of the Study:

• To develop a high-performance, nonvolatile memristor based on ferroelectric tunnel junctions for computing in memory (CIM).
• To demonstrate the device's capability in emulating synaptic functions for artificial intelligence applications.
• To address the limitations of the von Neumann architecture through novel memristor technology.

Main Methods:

• Fabrication of a Pd/Bi0.9La0.1FeO3 (BLFO)/La0.67Sr0.33MnO3 (LSMO) ferroelectric tunnel junction memristor on a silicon substrate.
• Utilizing ferroelectric polarization reversal triggered by pulse stimulation to modulate device conductance and achieve synaptic plasticity.
• Implementing the memristor for image processing tasks and simulating a VGG8 convolutional neural network (CNN).

Main Results:

• The FTJ memristor exhibited stable nonvolatile operation with tunable conductance states.
• The device successfully emulated various synaptic functions with high linearity and symmetry in weight updates.
• Demonstrated excellent performance in image processing and achieved 92.07% recognition accuracy for CNN offline learning on the Cifar-10 dataset.

Conclusions:

• The developed FTJ memristor provides a stable and high-performance solution for computing in memory applications.
• This technology offers a viable pathway to overcome the bottleneck of the von Neumann architecture.
• The device shows significant potential for future neuromorphic computing and AI hardware acceleration.