All-magnonic Neurons Enable Analog Neural Networks with Tunable Fading Memory and Multi-Input Integration

The pursuit of energy-efficient computing has driven significant interest in novel hardware approaches to artificial intelligence, and now, researchers are exploring the potential of wave-based systems. David Breitbach, Moritz Bechberger, and Hanadi Mortada, alongside colleagues at RPTU Kaiserslautern-Landau and the Institute of Magnetism of the NAS of Ukraine, demonstrate a fully functional artificial neuron built entirely from magnons, wave-like disturbances in magnetic materials. This all-magnonic neuron exhibits key characteristics of biological neurons, including a sharp response to input signals and a fading memory, and crucially, connects to other neurons via propagating magnons. The team’s work establishes these devices as promising building blocks for scalable, low-power neural networks, offering a potential pathway towards more efficient and powerful artificial intelligence systems.

Magnonic Neuromorphic Computing with Spin Waves

Scientists are exploring neuromorphic computing, building systems inspired by the brain’s structure and function, by utilizing magnonics and employing spin waves, known as magnons, as information carriers. This approach promises lower energy consumption, faster processing speeds, and brain-like computational architectures. The team investigates magnonic adders, neurons, and repeaters, built using gallium-substituted yttrium iron garnet, chosen for its low energy loss and efficient spin wave propagation. Researchers are harnessing nonlinear magnonics, exploiting effects that allow for complex functionalities like signal switching and frequency conversion.

This is achieved through manipulating spin wave formations, called bullets and droplets, and understanding wave turbulence, where energy cascades through a system of waves. Controlling this turbulence is crucial for creating stable devices. This work explores computational paradigms, including spiking neural networks and reservoir computing. Perturbation-based computing and Kolmogorov-Arnold Networks are also being investigated as alternatives to traditional deep learning. The team employs micro-focused Brillouin light scattering to image spin waves at the nanoscale, and ferromagnetic resonance to characterize their behavior. Thin film deposition and microwave techniques are used to create and detect spin waves. This interdisciplinary research presents a compelling vision for the future of computing, offering the potential for more energy-efficient and brain-inspired systems.

All-Magnonic Neurons Using YIG Thin Films

Scientists are developing all-magnonic neurons, devices utilizing magnons to overcome limitations of conventional digital neural networks. This work presents a novel approach to artificial neuron design, utilizing a gallium-substituted yttrium iron garnet film as the core component. Three nanofabricated coplanar waveguide antennas are positioned on this film, each functioning as an individual neuron, arranged in triangular and parallel series configurations to study neuron triggering and connectivity. The team engineered a nonlinear activation function by exploiting the nonlinear shift in magnon frequency with increasing intensity.

Magnons are excited by applying a radio frequency current, generating dynamic magnetic fields. Space- and time-resolved Brillouin light scattering spectroscopy is used to investigate the interaction of up to three magnonic neurons, revealing the nonlinear frequency shift caused by magnon-magnon interactions. This nonlinear response is crucial for achieving threshold-like behavior. By carefully controlling the excitation power and frequency, the team demonstrated intrinsic signal accumulation and threshold-triggered firing within a single neuron. This architecture inherently supports connectivity and cascadability without requiring external signal amplification. To validate functionality, the activation function was implemented within a simulated artificial neural network, achieving approximately 97% validation accuracy on the MNIST benchmark and 87% on fashionMNIST.

All-Magnonic Neurons Demonstrate Threshold Firing

Scientists have achieved a breakthrough in neuromorphic computing by realizing all-magnonic neurons based on a nonlinear excitation mechanism within gallium-substituted yttrium iron garnet films. This work demonstrates a pathway towards scalable, low-power artificial neural networks by harnessing the unique properties of magnons, experimentally verifying a threshold-like trigger response analogous to biological neuron activation. The team fabricated three nanofabricated coplanar waveguide antennas on a 56-nanometer gallium-substituted yttrium iron garnet film, each representing an individual neuron. Using space- and time-resolved Brillouin light scattering spectroscopy, they successfully demonstrated multi-neuron triggering, cascadability, and multi-input integration.

The system functions as both a nonlinear magnon emitter and amplifier, supporting connectivity without external amplification. Furthermore, the experimentally verified neuron activation function was implemented within a simulated artificial neural network to assess its suitability for large-scale neuromorphic systems. Despite experimental limitations, the network achieved validation accuracies of approximately 97% on the MNIST benchmark and 87% on fashionMNIST, establishing all-magnonic neurons as promising building blocks for future physical neural networks.

Magnonic Neurons Demonstrate Neuromorphic Computing Functions

This research demonstrates a novel approach to artificial neurons, utilizing magnons to perform computations. Scientists successfully created a fully magnonic neuron based on a gallium-substituted yttrium iron garnet film, achieving a tunable activation function through the nonlinear excitation of magnons, enabling a sharp response to input signals and incorporating a fading memory capability. The team established connections between these magnonic neurons via propagating magnons, experimentally linking up to three neurons and demonstrating essential functions for neuromorphic computing, including multi-input activation, signal integration, and cascading. Embedding a model of this experimentally derived neuron within a neural network simulation yielded high classification accuracy on standard benchmarks, validating its potential as a scalable building block for low-power, wave-based neural hardware. The authors acknowledge that the full temporal dynamics of the neuron were not explored in this study, representing a potential avenue for future research, particularly in the development of analog recurrent networks.