All-magnonic neurons for analog artificial neural networks

Abstract: Analog neuromorphic hardware is gaining traction as conventional digital systems struggle to keep pace with the growing energy and scalability demands of modern neural networks. Here, we present analog, fully magnonic, artificial neurons, which exploit a nonlinear magnon excitation mechanism based on the nonlinear magnonic frequency shift. This yields a sharp trigger response and tunable fading memory, as well as synaptic connections to other neurons via propagating magnons. Using micro-focused Brillouin light scattering spectroscopy on a Gallium-substituted yttrium iron garnet thin film, we show multi-neuron triggering, cascadability, and multi-input integration across interconnected neurons. Finally, we implement the experimentally verified neuron activation function in a neural network simulation, yielding high classification accuracy on standard benchmarks. The results establish all-magnonic neurons as promising devices for scalable, low-power, wave-based neuromorphic computing, highlighting their potential as building blocks for future physical neural networks.