Multiferroic antiferromagnetic artificial synapse

Abstract

Artificial intelligence frameworks utilizing unsupervised learning techniques can avoid the bottleneck of labeled training data required in supervised machine learning systems, but the programming time of these systems is inherently limited by their hardware implementations. Here, a finite-element model coupling micromagnetics and dynamic strain is used to investigate a multiferroic antiferromagnet as a high-speed artificial synapse in artificial intelligence applications. The stability of strain-induced intermediate antiferromagnetic magnetization states (non-uniform magnetization states between a uniform 0 or 1), along with the minimum time scale at which these states can be programmed is investigated. Results show that due to the antiferromagnetic material's magnetocrystalline anisotropy, two intermediate states (Néel vector 1/3z, 2/3x, and Néel vector 2/3z, 1/3x) between fully x and fully z Néel vector orientations can be successfully programmed using 375  με strain pulses, and that the time associated with this programming is limited to ∼0.3 ns by the material's antiferromagnetic resonance frequency.