Shape‐Dependent Multi‐Weight Magnetic Artificial Synapses for Neuromorphic Computing

Abstract

In neuromorphic computing, artificial synapses provide a multi‐weight (MW) conductance state that is set based on inputs from neurons, analogous to the brain. Herein, artificial synapses based on magnetic materials that use a magnetic tunnel junction (MTJ) and a magnetic domain wall (DW) are explored. By fabricating lithographic notches in a DW track underneath a single MTJ, 3–5 stable resistance states that can be repeatably controlled electrically using spin‐orbit torque are achieved. The effect of geometry on the synapse behavior is explored, showing that a trapezoidal device has asymmetric weight updates with high controllability, while a rectangular device has higher stochasticity, but with stable resistance levels. The device data is input into neuromorphic computing simulators to show the usefulness of application‐specific synaptic functions. Implementing an artificial neural network (NN) applied to streamed Fashion‐MNIST data, the trapezoidal magnetic synapse can be used as a metaplastic function for efficient online learning. Implementing a convolutional NN for CIFAR‐100 image recognition, the rectangular magnetic synapse achieves near‐ideal inference accuracy, due to the stability of its resistance levels. This work shows MW magnetic synapses are a feasible technology for neuromorphic computing and provides design guidelines for emerging artificial synapse technologies.

Demonstration of Synaptic Behavior in a Heavy-Metal-Ferromagnetic-Metal-Oxide-Heterostructure-Based Spintronic Device for On-Chip Learning in Crossbar-Array-Based Neural Networks

“…As mentioned in the previous section, the average energy per LTP/LTD step in our projected spintronic synapse device is 150 fJ. This value is comparable to that reported for the spintronic synapse by Leonard et al The report by Leonard et al also compares this energy value with that reported for other NVM synapse technologies (RRAM-based and PCM-based) elsewhere, which are mostly in nanoJoules and picoJoules. Thus, like the device in Leonard et al, our projected spintronic synapse device can consume orders of magnitude (1–4 orders) lower energy than other NVM synapse devices.…”

“…S. Zhang et al and Xie et al do not carry out any such stability study. Leonard et al show only 3–5 stable states per device (so only 1–2 bits), but those states are not obtained through identical-magnitude “write” pulses like we have here (this makes the peripheral circuit design simpler for us compared to that by Leonard et al) . Q. Zhang et al also use nonidentical pulses (pulses of increasing current magnitude) for LTP and LTD .…”

“…Once an MTJ is fabricated from the stack, these states will correspond to different synaptic conductance states because of tunneling magneto-resistance (TMR). While such mixed states have been reported in Ta/CoFeB/MgO heterostructures ,, and CoPt alloy films, to the best of our knowledge, mixed states have not been reported in Pt/Co/SiO 2 heterostructures before this. We experimentally demonstrate control of the mixed states with identical programming current, or “write” current pulses. We obtain gradual increase (LTP) and gradual decrease (LTD) of anomalous Hall resistance within a certain regime.…”

“…This is because for these RRAM devices, to obtain desirable LTP and LTD characteristics, all LTP pulse units cannot be identical to each other and all LTD pulse units cannot be identical to each other. 30,71−73 The programming pulses used are nonidentical for spintronic devices as well in some reports published elsewhere, 11,43 as we have mentioned in the Introduction section. This is not the case with our spintronic device, as we just mentioned above.…”

“…Nanomagnetic and spintronic devices based on the heavy-metal-ferromagnetic-metal-oxide heterostructures have been considered very attractive for various applications including nonvolatile memory (NVM), − logic, and neuromorphic computing. − These devices heavily use interfacial phenomena like perpendicular magnetic anisotropy (PMA) and spin–orbit torque (SOT) for their functionalities. − ,− PMA is considered to originate primarily from the ferromagnet–oxide interface, , while the following mechanisms have been considered as origins of SOT: Spin Hall effect inside the heavy metal leads to accumulation of spin-polarized electrons at the heavy-metal–ferromagnet interface; these spin-polarized electrons subsequently flow from the heavy metal to the ferromagnet to apply SOT on the magnetic moments of the ferromagnetic layer. ,, Rashba effect at the heavy-metal–ferromagnet interface Additional interface-originated effects like the orbital Rashba–Edelstein effect, recently reported by Chen et al …”