Field-Free Spin-Orbit Torque Devices for Neural Network Applications

Robust field-free spin-orbit torque (SOT) switching plays an essential role in realizing practical SOT-based magnetic memories and computing devices. This work explores the use of antisymmetric interlayer Dzyaloshinskii–Moriya interaction (i-DMI), which facilitates the emergence of noncollinear chiral spin configurations with wedge deposition. The strong i-DMI is obtained in an optimized CoFeB/Pt(thick)/Co/Pt(thin) heterostructure which is enhanced by tuning of the thickness of a capping layer due to the interference of the electron wave function. Herein field-free SOT switching is achieved with a threshold switching current density of 5.4 × 10¹⁰ A/m². For neural network applications, SOT devices enable both analog and digital synapses, with a single analog cell or multiple binary cells simulating multilevel weights. This work demonstrates the proof of concept by using Hall bar devices to mimic memristive behaviors, applying consecutive current pulses to a single Hall bar and separate current pulses to a quad-channel Hall bar device. Our simulation results show that using 2-bit digital SOT MRAM achieves a classification accuracy of 92.2% on the MNIST dataset, significantly outperforming the analog SOT MRAM case (<10%) and demonstrating remarkable improvements in power and time efficiencies compared to the SRAM case. These results highlight digital SOT MRAM as a top choice for edge AI and offline inference applications.