Reconfigurable Training and Reservoir Computing with Multi-State Artificial Spin Ice Networks

The key technological challenge is achieving high-speed, accurate, and energy-efficient information processing. While conventional machine learning excels at prediction, reservoir computing (RC) offers a resource-efficient alternative, requiring minimal training data, relying on linear optimization, and operating at low computational cost [1-3]. Here, we use elongated Ni₈₁Fe₁₉ nanodots arranged in a square artificial spin ice (ASI) for reconfigurable training and reservoir computing. We train three different combinations of magnetization configurations (macrospin, single-vortex, and double-vortex states) with various pulse waveforms—sawtooth, square, triangular, sine, and Mackey–Glass. The RC framework comprises an input layer, the ASI processing layer that hosts GHz-range resonance frequencies determined by microstate fingerprints, and an output layer applying ridge regression with nonlinear transformation to generate predictions. Our results demonstrate that the used square ASI supports macrospin, single-vortex, and double-vortex-based training—similar to binary spin-vortex ice [1]—underscoring its effectiveness as a platform for reservoir computing.

References:

[1] J. C. Gartside, K. D. Stenning, A. Vanstone, H. H. Holder, et al. Nat. Nanotechnol. 17, 460 (2022).

[2] W. Hu, Z. Zhang, Y. Liao, Q. Li, et al. Nat. Commun. 14, 2562 (2023).

[3] R. Sultana, A. K. Mondal, V. S. Bhat, K. Stenning, et al. J. Appl. Phys. 138, 061101 (2025).