Ice Sculpting: Neuromorphic Training of Geometrically-Frustrated Magnetic Metamaterials

Artificial spin ice (ASI) consists of ferromagnetic arrays populated by Ising-like nanopatterned macrospins. The array ‘microstate’ describes the magnetic orientations of all macrospins, with the ASI microstate manifold notable for its vast range of energetically-degenerate states arising from geometric frustration.
ASI has displayed capacity to ‘memorise’ specific microstates, returning to them with perfect fidelity even under strong magnetic perturbations. This property, termed return point memory (RPM), is a signature of systems which may be trained to act as artificial neural networks.
Today’s software-based neural networks are powerful but crucially limited by their non-neuromorphic host hardware. Designing novel computational hardware which is fundamentally neuromorphic in design and operation may greatly enhance the scope and utility of artificial neural networks.
Combining recent development of a nanomagnetic writing technique allowing for access to the entire ASI microstate space and computer-vision microstate recognition with the inherent RPM properties of ASI, we explore the viability of ASI as a hardware platform for directly implementing neural net functionality with no software layer.