Multilayer DW & Skyrmion-based Spintronic LIF Neuron Model & Devices for Neuromorphic Computing

What is it about?

Neuromorphic computing, inspired by the structure and function of the human brain, offers groundbreaking potential for energy-efficient and adaptive computing systems. Unlike traditional computing models, neuromorphic systems operate through spiking neural networks (SNNs), which mimic the dynamic and parallel processing capabilities of biological neurons. These systems excel in real-time data processing, scalability, and robustness, making them ideal candidates for a wide range of applications including perception, image classification, pattern recognition, and machine learning. Traditional implementations of SNNs often rely on CMOS-based hardware, such as IBM's TrueNorth and Intel's Loihi chips, to simulate neural dynamics. However, these devices face significant limitations in terms of energy consumption, scalability, and speed. This study proposes spintronic based neuromorphic devices using domain walls and skyrmions in multilayer ferromagnetic structures to emulate leaky integrate-and-fire neurons. By combining spin–orbit torque, demagnetization energy modulation, and micromagnetic simulations, these devices achieve efficient spike latency, integration density and robust LIF behavior. Integrated into spiking neural networks, they achieve high-density computing, 98% accuracy on MNIST, and 95% on FMNIST, paving the way for scalable, energy-efficient neuromorphic hardware.

Featured Image

Photo by Bhautik Patel on Unsplash

Photo by Bhautik Patel on Unsplash

Why is it important?

The key aspects of the proposal are that the proposed neuron model incorporates spintronic mechanisms, and the LIF (leaky integrate-and-fire) model is achieved through the interaction of spin-transfer torque (STT), spin-orbit torque (SOT), demagnetization energy, and external input pulses. This enables precise control over neural spike timing and inter-neuron communication. The proposed structure is based on multilayer ferromagnetic materials, which host domain walls and skyrmions at ultra-small sizes—critical for achieving high-density integration and scalability. These two challenges have been major obstacles in traditional neuromorphic computing architectures. Unlike single-layer structures, the multilayer configuration enhances the stability of magnetic solitons, improving their resilience to thermal fluctuations and external perturbations. This increased stability is essential for long-term operation and performance, allowing the integration of a larger number of devices without sacrificing reliability. Previous experimental studies have demonstrated that multilayer systems, such as CoFeB/Ta, exhibit enhanced stability of domain walls and skyrmions, making them ideal for use in high-density neural networks. Furthermore, the multilayer design ensures that these devices can be scaled down effectively, enabling the integration of vast numbers of neurons and synapses within neuromorphic systems. In particular, skyrmion-based devices require low driving currents for propagation, which results in ultra-low power consumption. Additionally, the latency typically associated with skyrmion motion—a challenge for real-time applications—is mitigated through careful engineering of the multilayer structure and SOT dynamics. This leads to faster response times and improved system efficiency. Another significant aspect of this research is the potential for seamless integration with CMOS technologies. Spintronic devices, such as domain wall magnetoresistive tunnel junctions (DW-MTJs), exhibit a linear relationship between their magnetoresistance and external stimuli. This property makes them ideal candidates for implementing both linear weight updating and non-linear activation functions needed in spiking neural networks (SNNs). The ability to integrate these spintronic devices with existing CMOS structures allows for a smooth transition to spintronic neuromorphic systems, without requiring a complete overhaul of current technologies. Micromagnetic simulations effectively capture the behavior of the LIF neurons under applied current pulses, confirming that the proposed devices accurately model the dynamics of spiking neurons. Furthermore, the integration of these neurons into a three-layer SNN and a convolutional SNN framework demonstrates their practical applicability in machine learning tasks. The network achieves impressive performance, with classification accuracy exceeding 96% for the MNIST and FMNIST datasets, illustrating the feasibility of spintronic-based SNNs for high-accuracy pattern recognition.

Perspectives