Artificial Neurons Ditch Magnetic Fields for More Powerful, Scalable Computing

Researchers are developing artificial neurons using spintronics to create more energy-efficient computing systems. Badsha Sekh, Hasibur Rahaman, and Ravi Shankar Verma, from Nanyang Technological University and the Indian Institute of Technology Roorkee, alongside Ramu Maddu, Kesavan Jawahar, and S. N. Piramanayagam, have demonstrated a novel spintronic neuron that overcomes a key limitation of current designs. Their work eliminates the need for external magnetic fields by utilising out-of-plane spin-polarised spin-splitting torque and incorporating a self-reset function via synthetic antiferromagnetic coupling. Significantly, this device achieved high accuracy, 95.99% on the MNIST dataset and 94.36% on N-MNIST, validating its potential for building compact, scalable, and truly energy-efficient neural networks.

Ruthenium dioxide altermagnetism enables self-resetting spintronic neuron operation with high energy efficiency

Scientists have developed a novel spintronic artificial neuron that eliminates the need for external magnetic fields, a common limitation in contemporary designs. This breakthrough, achieved through an altermagnet/Synthetic Antiferromagnetic Coupling (SAF) based device, utilizes out-of-plane spin-polarized spin-splitting torque to facilitate neuron operation.
The research addresses a critical challenge in neuromorphic computing by creating a neuron with an intrinsic self-reset function, enabled by built-in exchange coupling, thereby simplifying system architecture and reducing power consumption. This innovative approach leverages the unique properties of altermagnets, specifically ruthenium dioxide, to generate a spin current that drives deterministic magnetization switching without requiring an external in-plane magnetic field.

The device integrates ruthenium dioxide with a platinum spin-orbit torque layer, combining two distinct mechanisms to achieve efficient magnetization switching at zero applied field. Application of an in-plane charge current generates spin-orbit torques from the platinum layer and, crucially, a crystal angle dependent, out-of-plane polarized spin-splitting torque from the ruthenium dioxide.
These combined torques overcome the antiferromagnetic interlayer exchange coupling of the SAF stack, inducing magnetization reversal in the soft layer, which represents the neuron’s integration activity. Upon removal of the current, the exchange coupling automatically resets the magnetization, mimicking the behaviour of biological neurons.

Experimentally, researchers epitaxially grew an 18nm ruthenium dioxide thin film on an aluminium oxide substrate, confirmed by Θ-2Θ XRD data. A series of SAF stacks were then deposited, varying layer thicknesses to optimise the exchange coupling and minimize the magnetization switching field. MOKE hysteresis loops demonstrated sharp perpendicular magnetic anisotropy across the sample series, confirming robust magnetic properties. Validation of the proposed spintronic neuron for Spiking Neural Network (SNN) applications yielded test accuracies of 95.99% on the MNIST dataset and 94.36% on the N-MNIST dataset, demonstrating hardware feasibility and compatibility for compact, scalable, and energy-efficient neuromorphic systems.

Altermagnet SAF neuron fabrication and spiking neural network implementation represent a promising avenue for neuromorphic computing

An altermagnet/Synthetic Antiferromagnetic Coupling (SAF) based spintronic neuron forms the core of this research, designed to operate without requiring external in-plane magnetic fields. The device leverages out-of-plane spin (σ_Z) polarized spin-splitting torque to achieve neuron functionality, eliminating the need for conventional symmetry-breaking fields that hinder scalability.

Fabrication involved creating a heterostructure incorporating the altermagnet and SAF layers, enabling intrinsic self-resetting behaviour through built-in exchange coupling. To validate the neuron’s performance, researchers integrated it into a Spiking Neural Network (SNN) and tested its capabilities on the MNIST and N-MNIST datasets.

N-MNIST event streams were discretized into millisecond-scale time bins, accumulating events within each bin to create temporal frames for sequential processing by the SNN. The network employed adaptive Leaky Integrate-and-Fire (LIF) neurons with dual time constants (τ), trained using a supervised surrogate-gradient backpropagation method.

Classification accuracy was determined by the cumulative spike count of output neurons, identifying the predicted class as the neuron exhibiting the highest spiking activity during inference. This approach yielded test accuracies of 95.99% on MNIST and 94.36% on N-MNIST, demonstrating competitive performance compared to existing supervised and unsupervised SNN models.

The study also details a confusion matrix analysis, revealing class-wise prediction behaviour and misclassification patterns, confirming the model’s reliable digit classification capabilities. This methodology establishes the proposed spintronic neuron as a viable platform for energy-efficient neuromorphic computing.

High accuracy spintronic neuron operation via out-of-plane torque and self-resetting functionality is demonstrated

Researchers achieved test accuracies of 95.99% and 94.36% on the MNIST and N-MNIST datasets, respectively, using a novel spintronic neuron design. This work demonstrates a functional Spiking Neural Network (SNN) utilising an altermagnet/Synthetic Antiferromagnetic Coupling (SAF) based neuron, eliminating the need for external in-plane magnetic fields.

The device operates by leveraging out-of-plane spin (σZ) polarized spin-splitting torque, facilitating neuron operation without external assistance. The fabricated neuron features an intrinsic self-reset function, achieved through built-in exchange coupling within the SAF stack. Table 1 details the full SAF stack, with layer thicknesses specified in nanometers, enabling optimisation of the interlayer exchange field.

Epitaxial growth of the RuO2 (101) thin film was confirmed via Θ-2Θ XRD data, demonstrating the quality of the material used in device fabrication. During experimentation, application of in-plane charge current (JC) generated spin-orbit torques from both a Pt layer and the RuO2 altermagnet. The combined torques compete against the Hex of the SAF, inducing deterministic magnetization reversal of the soft layer when exceeding the switching field.

This magnetization reversal emulates the integration activity of the neuron, while the Hex facilitates restoration to the initial magnetic configuration, mimicking the reset functionality. The study successfully demonstrates hardware feasibility and compatibility for compact, scalable, and energy-efficient neuromorphic computing systems.

Spintronic neuron achieves high accuracy in benchmark neural network tests, demonstrating promising potential

Researchers have demonstrated a spintronic neuron based on an altermagnet and synthetic antiferromagnetic coupling, eliminating the need for external magnetic fields during operation. This device utilises out-of-plane spin-polarised spin-splitting torque to achieve neuronal function, offering a pathway towards more scalable and practical neuromorphic computing systems.

The neuron also incorporates an intrinsic self-reset function, facilitated by built-in exchange coupling, simplifying its design and operation. Validation of the proposed device was performed through spiking neural network applications, achieving test accuracies of 95.99% and 94.36% on the MNIST and N-MNIST datasets, respectively.

These results confirm the hardware feasibility and compatibility of the spintronic neuron for use in complex neural networks. The authors acknowledge that the performance reduction observed on the N-MNIST dataset, compared to MNIST, aligns with findings from previous neuromorphic studies. Future research could focus on optimising the device’s performance further and exploring its integration into larger-scale neuromorphic architectures.

This work establishes a viable and scalable platform for compact, energy-efficient, and external field-free spintronic neurons, representing a significant step towards practical neuromorphic hardware integration. The demonstrated self-reset mechanism and compatibility with standard datasets highlight the potential of this technology for event-driven vision devices and other low-power applications.