Radiofrequency Spintronic Neural Network Enabled by Electrically Modulated Magnetic Tunnel Junctions.

Area of Science:

• Spintronics
• Neuromorphic Computing
• Materials Science

Background:

• Magnetic tunnel junctions (MTJs) show promise for energy-efficient neuromorphic computing.
• Existing MTJ-based systems often use magnetic fields, leading to high energy use and limited scalability.
• Limited tunable bandwidth in MTJs restricts synapse count and network scaling.

Purpose of the Study:

• To experimentally realize electrically tunable spintronic synapses and neurons using three-terminal MTJs.
• To demonstrate a scalable and energy-efficient neuromorphic computing system.
• To explore the potential of spintronic devices for advanced AI applications.

Main Methods:

• Utilized three-terminal MTJs for spintronic synapses and neurons.
• Employed spin-orbit torque for precise modulation of synaptic weight and neuron output frequency.
• Stimulated multilayer networks with fully connected and convolutional architectures.

Main Results:

• Achieved precise electrical control over synaptic weight and neuron output frequency.
• Reduced energy consumption by a factor of 21 compared to existing methods.
• Demonstrated high accuracy (99.2% on drone classification, 92.0% on Fashion-MNIST) in multilayer networks.
• Convolutional design reduced the need for oscillator frequency channels.

Conclusions:

• Successfully demonstrated scalable, high-frequency, and energy-efficient all-spintronic neuromorphic systems.
• The proposed method offers a compatible platform for future neuromorphic computing.
• Electrical tunability overcomes limitations of magnetic field modulation for MTJ-based neural networks.