Magnetic skyrmions, which are topological magnetic solitons, boast a range of enticing features that make them promising candidates for energy-efficient computing operations [1-3], such as stability at room temperature, deep sub-micron dimensions, non-volatility, particle-like behavior, and motion at low power. These characteristics seamlessly align with the requisites of neuromorphic computing, rendering them attractive candidates for integration into neuromorphic circuits [1-5]. Recent experimental studies have shown that magnetic skyrmions can be nucleated [6-9], moved [7,9], annihilated [8] and electrically detected using the anomalous Hall effect [10] or tunneling magnetoresistance [11-12]. However, an experimental demonstration of the weighted sum operation is still missing in the context of skyrmions [4]. In our recent works [13-14], we leverage the non-volatile and particle-like characteristics of magnetic skyrmions, akin to synaptic vesicles containing neurotransmitters, to perform this weighted sum operation in a compact, biologically-inspired manner (see Fig. 1a-b). Our experimental proof of concept demonstrates a large-scale, low-energy hardware implementation of weighted sum operations based on magnetic skyrmions (see Fig. 1c-g) [13]. We demonstrate the precise electrical control of skyrmion nucleation and motion in specially designed magnetic tracks, the number of generated skyrmions being determined by the electrical pulse input multiplied by the track synaptic weight. Leveraging magneto-ionic effects to achieve non-volatile and reversible control over magnetic properties opens the door to gate voltage control of synaptic weights. Voltage gating control of magnetic anisotropy is achieved through the application of an electric field through an AlO x layer, enabling the in-plane to out-of-plane magnetic anisotropy switch in the top Co layer [14]. Detection of the number of skyrmions is accomplished through non-perturbative anomalous Hall voltage measurements. We experimentally validate the weighted sum operation using two electrical inputs in a crossbar array configuration with two tracks (see Fig. 1c-g) [13]. This ensures efficient execution of the fundamental weighted sum operation, a cornerstone for neuromorphic computing. Our experimental demonstration is scalable to accommodate multiple inputs and outputs using a crossbar array design, potentially approaching the energy efficiency observed in biological systems.

[1] A. Fert, N. Reyren, and V. Cros, Nat Rev Mater 2, 17031, 2017.
[2] G. Bourianoff et al., AIP Advances, vol. 8(5), pp. 055602, 2018.
[3] K.M. Song et al., Nature Electronics, vol. 3.3, pp. 148-155, 2020.v [4] J. Grollier et al., Nature Electronics 3, 360–370, 2020.
[5] Y. Huang et al., Nanotechnology 28, 08LT02 (2017).
[6] W. Legrand et al., Nano Letters, vol. 17, pp. 2703, 2017.
[7] A. Hrabec et al., Nature Communications, vol. 8, pp. 1-6, 2017.
[8] S. Woo et al., Nature Electronics, vol. 1, pp. 288-296, 2018.
[9] S. Woo et al., Nature Materials, vol. 15, pp. 501-506, 2016.
[10] D. Maccariello et al., Nature Nanotechnology, vol. 13, pp. 233-237, 2018.
[11] C. Hanneken et al., Nature Nanotechnology 10, 1039–1042, 2015.v [12] J. U. Larrañaga et al., arXiv:2308.00445, 2023.
[13] T. da Câmara Santa Clara Gomes et al., arXiv:2310.16909, 2023.
[14] T. da Câmara Santa Clara Gomes et al., arXiv:2310.01623, 2023.


​ Plus d'information :https://www.spintec.fr/seminar-neuromorphic-weighted-sum-with-magnetic-skyrmions/​​​