Shape Anisotropy-Dependent Leaking in Magnetic Neurons for Bio-Mimetic Neuromorphic Computing

Area of Science:

• Neuromorphic Engineering
• Spintronics
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) aim to replicate biological computation using artificial neurons and synapses.
• Spintronic devices offer a path to emulate neuron functions like integration and firing using magnetic phenomena.
• Leaky integrate-and-fire (LIF) behavior, crucial for neuron relaxation, remains underexplored in spintronic prototypes.

Purpose of the Study:

• To investigate domain wall-magnetic tunnel junction (DW-MTJ) devices for artificial neurons capable of LIF behavior.
• To demonstrate and analyze geometry-dependent leaking dynamics for tunable LIF operation.
• To explore methods for enhancing neuron expressivity without compromising leaking dynamics.

Main Methods:

• Fabrication and characterization of five distinct DW-MTJ device designs.
• Systematic tuning of device geometry, stimulation fields/currents, and contact placement.
• Implementation of asymmetric notches to induce nonlinear pinning.
• Simulation of a spiking neural network using measured DW-MTJ neuron behavior.

Main Results:

• Demonstrated geometry-dependent leaking dynamics enabling repeatable and tunable LIF operation in DW-MTJ neurons.
• Showcased a wide range of neuron behaviors by manipulating device geometry, stimuli, and contact locations.
• Introduced nonlinear pinning via asymmetric notches, increasing neuron expressivity while preserving leaking.
• Simulated SNNs with DW-MTJ neurons outperformed 1D continuous DW motion models.

Conclusions:

• DW-MTJ artificial neurons exhibit analog LIF capabilities, integrating multiple desirable neuron functions into a single device.
• The demonstrated tunable LIF behavior and enhanced expressivity pave the way for multifunctional neuromorphic computing.
• This work advances the development of efficient and versatile spintronic neuromorphic hardware.