Low-Power Artificial Neural Network Perceptron Based on Monolayer MoS₂

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Edge computing and the Internet of Things (IoT) require highly energy-efficient processors for data-driven applications.
• Traditional silicon-based processors face limitations in energy efficiency for complex machine learning tasks.
• Advancements in two-dimensional (2D) materials offer potential for novel electronic device architectures.

Purpose of the Study:

• To demonstrate an in-memory processor utilizing a 2D materials platform for enhanced artificial neural network (ANN) performance.
• To evaluate the scalability and memory performance of 2D channel flash memory devices.
• To assess the potential of these devices to outperform silicon counterparts in energy efficiency and computational capability.

Main Methods:

• Fabrication of a flash memory array with a two-dimensional channel using wafer-scale Molybdenum Disulfide (MoS2).
• Device characterization through simulations and experiments to assess scalability and memory performance.
• Closed-loop programming for precise tuning of device conductance (4-bit precision).
• Demonstration of a physical circuit for seven-segment digit classification.

Main Results:

• The 2D MoS2 flash memory array demonstrates scalability to sub-micrometer channel lengths with sustained memory performance.
• Simulations indicate a viable memory window even at sub-50 nm channel lengths.
• A physical circuit achieved 91.5% accuracy in seven-segment digit classification with ex situ training.
• System-level simulations project potential energy efficiency of 50,000 TOPS/W for AlexNet classification.

Conclusions:

• In-memory processors based on 2D materials, specifically MoS2, offer a promising alternative to silicon for energy-efficient ANNs.
• The developed flash memory technology is scalable and maintains performance at reduced dimensions.
• This platform has the potential to significantly advance edge computing and IoT applications by enabling more powerful and efficient on-device AI.