Multifunctional vortex magnetic tunnel junctions with optical and radio-frequency inputs for neuromorphic computing

Abstract The brain co-locates memory and processing, but many neuromorphic devices need standby power or separate memory from nonlinear computation. Here we show that vortex magnetic tunnel junctions with a non-volatile, thermally programmable storage layer provide two complementary nonlinear channels in one nanopillar. The laser-induced voltage is linear in laser power but cubic in direct-current bias, making the bias current a nonlinear, programmable weight, whereas radio-frequency excitation is rectified into a frequency-dependent voltage. Crossbar simulations using measured responses classify handwritten digits with 95.4\% accuracy using optical inputs and 94.9\% using radio-frequency inputs, comparable to an equal-size digital single-layer network. Microlenses focusing light to sub-micrometre spots, one formed directly on a junction, and a flip-chip-mounted vertical-cavity surface-emitting laser array provide a route to on-chip optical addressing. Operating both channels together could combine optical and radio-frequency inputs within one nanoscale device, making vortex junctions reprogrammable multi-input building blocks for analog neuromorphic hardware.