Strain Controlled Domain Wall Synapse with Quantized Weights in the Presence of Thermal Noise and Edge Roughness

We perform micromagnetic simulations to show that energy efficient [1] strain control of domain wall (DW) in a perpendicularly magnetized racetrack can realize a multi-state synapse well suited to neuromorphic computing-based classification tasks. In conjunction with a fixed current exerting spin orbit torque (SOT), the racetrack is strained to modulate its anisotropy by applying voltage across the piezoelectric substrate on which the racetrack is patterned to control the translation of the DW to different positions in the racetrack. Simulations that include edge roughness and thermal noise showed that 5-state and 3-state synapse are possible in a 500 nm long and 50 nm wide racetrack [2]. Such limited state DW synapse is attractive to implement quantized neural network which is proven to achieve near equivalent accuracy to full-precision network [3] even in the presence of device variability [4]. Preliminary experiments with such racetracks will also be presented.

[1]. M. A. Azam et al., Nanotechnology (2020)
[2]. W. A. Misba et al., https://arxiv.org/abs/2010.10076
[3]. I. Hubara et al., J. Mach. Learn. Res (2017)
[4]. V. Joshi et al., Nat. Commun. (2020)