Nonlinear gyrotropic magnetic vortices for efficient pattern recognition

Abstract

In nature-inspired computing, the search for compact and efficient nonlinear elements is key to mimicking the brain's remarkable capabilities. Spintronic oscillators offer a promising platform due to their inherent nonlinearity and complementary metal–oxide–semiconductor compatibility. However, the practical deployment of such devices remains limited by their dependence on external magnetic fields. This limitation can be addressed using nanoscopic magnetic vortices in spin-torque vortex oscillators (STVOs), which enable low-power field-free operation along with nanoscale dimensions, structural robustness, and low excitation thresholds. In this work, we exploit the short-term plasticity of magnetic vortices to demonstrate that the gyrotropic motion of the magnetic vortex in STVOs can be nonlinearly transformed by harnessing input pulse streams and can be utilized for classification tasks in a field-free condition. This nonlinear gyrotropic motion assists in real-time feature extraction and classification of multibit input patterns. We evaluate the performance of STVOs in classifying 4-bit and 6-bit digital input patterns using a feedforward neural network trained on the Modified National Institute of Standards and Technology handwritten image dataset. The system achieves classification accuracies of 90.1% for 4-bit and 88.3% for 6-bit inputs, while significantly reducing the synaptic weight count by 99%, compared to conventional software-based artificial neural networks, which can be beneficial for faster inference with reduced computations.