Leveraging Local Magnetic Modifications for Spin-Based Neuromorphic Computing.

Artificial intelligence applications require energy-efficient computing architectures beyond conventional von Neumann systems. Neuromorphic computing (NC) has emerged as a promising alternative, where spintronic domain wall (DW) devices are particularly attractive owing to their low-energy operation, non-volatility, and high endurance. In this work, we investigate spin-orbit torque (SOT)-driven DW synaptic devices with locally engineered interfacial Dzyaloshinskii-Moriya interaction (iDMI) and perpendicular magnetic anisotropy (PMA) modifications for multistate neuromorphic applications. Systematic studies reveal that the DW pinning strength and propagation dynamics are strongly governed by the geometrical parameters of the engineered pinning sites, particularly their length (Lp) and etching depth (Detch), enabling controllable tuning of the local energy landscape and depinning behavior. Optimized devices (D7; Lp = 5 μm, Detch = 1 nm) containing five engineered pinning sites exhibit deterministic staircase-like DW propagation with seven stable intermediate magnetization states suitable for synaptic operation. Micromagnetic simulations further confirm that localized iDMI and PMA modifications can generate robust pinning/depinning dynamics. Using experimentally derived multistate quantization levels, artificial neural network simulations implemented in PyTorch achieved image-classification accuracies of approximately 97% and 84% on the MNIST and Fashion-MNIST datasets, respectively. These results establish engineered SOT-DW devices as a promising platform for scalable and energy-efficient spintronic NC.