This paper demonstrates the design of a neural network hardware where all neuromorphic computing functions, including signal routing and nonlinear activation, are performed by spin-wave propagation and interference. Weights and interconnections are realized by a magnetic-field pattern applied to the substrate, scattering the spin waves. A custom micromagnetic solver, based on PyTorch, inverse-designs the scatterer. The study shows that spin waves transition from linear to nonlinear interference at high intensities, greatly increasing computational power in the nonlinear regime. The authors envision small-scale, compact, and low-power neural networks operating entirely within the spin-wave domain.

Ádám Papp, Wolfgang Porod, Gyorgy Csaba