Large-scale crossbar arrays based on three-terminal MoS₂ memtransistors.

Area of Science:

• Materials Science
• Artificial Intelligence
• Device Engineering

Background:

• Memristive crossbar architectures are key for edge AI but struggle with similar inputs.
• Existing solutions for output ambiguity are often impractical for resource-limited edge devices.

Purpose of the Study:

• To develop and demonstrate large-scale MoS2 memtransistor crossbar arrays for efficient edge AI inference.
• To show that gate control in memtransistors can resolve inference ambiguities without retraining.

Main Methods:

• Fabrication of dense, large-scale crossbar arrays with up to 2048 MoS2 memtransistors.
• Characterization of memtransistor performance: yield, write energy, read margins, and retention.
• Validation of array performance using MNIST handwritten digit classification.

Main Results:

• Achieved >92% yield across multiple large-scale arrays.
• Demonstrated low write energies (~0.2 fJ), high read margins (10^5), and projected retention >3 years.
• Successfully resolved inference ambiguities via gate modulation and classified MNIST digits.

Conclusions:

• MoS2 memtransistor crossbar architectures offer a viable solution for resolving inference ambiguities in edge AI.
• Gate modulation provides an effective, retraining-free method for enhancing separability.
• This work advances in-memory computing for decentralized AI applications.