Spin-splitting-torque-driven field-free perpendicular magnetization switching in RuO2/synthetic antiferromagnet heterostructures for spintronic convolutional neural networks

With the growing demand for low-power and high-speed spintronic devices, the development of advanced material systems with efficient spin control capabilities has emerged as a central focus in spintronics research. Here, we propose a fully antiferromagnetic device architecture based on a magnetically compensated RuO2/synthetic antiferromagnet heterostructure, achieving fully electrical writing and reading functionalities. This design, characterized by its negligible stray field and deterministic field-free switching, is inherently suitable for large-scale neuromorphic integration. In a proof-of-concept demonstration, we showcase the implementation of an all-spintronic convolutional neural network using this architecture, achieving a high recognition accuracy of 98.7% on the handwritten digit classification task.