A Floating-Gate Photoelectric Synaptic Transistor Utilizing BP/PO_x/WSe₂ Heterostructure for Neuromorphic Visual...

Area of Science:

• Materials Science
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Traditional AI hardware faces limitations in energy efficiency and scalability due to von Neumann architectures and silicon-based synapses.
• Two-dimensional (2D) van der Waals (vdW) materials offer unique properties like atomic thickness and tunable electronics for advanced hardware.
• Developing efficient synaptic devices is crucial for next-generation neuromorphic computing.

Purpose of the Study:

• To engineer a 2D van der Waals (vdW) heterostructure transistor for efficient neuromorphic hardware.
• To demonstrate synaptic functionalities and human visual memory mimicry in the device.
• To integrate the device into a convolutional neural network (CNN) for advanced AI tasks.

Main Methods:

• Fabrication of a vdW floating-gate transistor using Black Phosphorus (BP) and Tungsten Diselenide (WSe2) with a P-oxide (POx) insulator.
• Characterization of the transistor's electrical properties, including on-off current ratio and memory window.
• Demonstration of key synaptic plasticity behaviors (STP, LTP, PPF, LTP/D) and visual memory emulation.
• Implementation of a two-path CNN utilizing both optical and electronic inputs for face recognition.

Main Results:

• The vdW transistor achieved a high on-off current ratio (≈10^5) and a large memory window (73 V).
• The device successfully mimicked human visual memory under optical stimuli with ultralow energy consumption (10 pJ/event).
• A CNN incorporating the device achieved 96.9% accuracy on the Labeled Faces in the Wild (LFW) dataset.
• The heterostructure engineering optimized interface band alignment for enhanced performance.

Conclusions:

• The developed vdW transistor represents a significant advancement in neuromorphic hardware, overcoming limitations of current AI architectures.
• This technology enables energy-efficient, human-like visual processing for edge AI applications.
• The synergistic integration of optical and electronic inputs in the CNN showcases a promising direction for future AI systems.