Oxygen-Migration-Based Spintronic Device Emulating a Biological Synapse

Electronic emulation of the biological synapse, which is the memory and learning element of the brain, is an important step towards the realization of brain-inspired computing systems. However, a complementary metal-oxide-semiconductor−based implementation of a synapse is not scalable due to the power and area inefficiency. It is therefore essential to develop alternative material and device concepts that can mimic the maximum number of synaptic functionalities in an energy-efficient way. Here, we demonstrate a nanosized energy-efficient three-terminal artificial synapse, with a separate read and write path, based on a magnetic material ($\mathrm{Co}$) that could generate both positive and negative synaptic outputs. The $\mathrm{Co}$ magnetization, which is representative of the synaptic weight, is modulated in an energy-efficient way by using an electric field to transport oxygen ions in and out of the $\mathrm{Co}$ layer. A wide range of synaptic functions such as synaptic potentiation and depression, spike magnitude, rate, and timing-dependent plasticity, and the transition from short-term to long-term plasticity are demonstrated. Our results suggest the viability of a spintronic synapse in neuromorphic computing.

Figure 1

(a) Schematic of the film structure. (b) Schematic of the Hall bar before electrode deposition. Enlarged view of the Hall cross is shown in the inset. (c) State of the magnetic synapse under the application of negative gate voltage (V_g). Blue spheres represent oxygen ions. (d) State of the magnetic synapse under the application of positive gate voltage.

Figure 2

(a) Potentiation and depression of the magnetic synapse using positive and negative gate-voltage pulses. Magnetic hysteresis loops for the magnetic state with minimum and maximum magnetic weights are also shown. The $\mathrm{Co}$ magnetization is switched between the up and down state to emulate positive and negative synaptic weight. (b) Dependence of the anomalous Hall resistance (R_AHE) or the synaptic weight on the magnitude of the V_g pulse.

Figure 3

(a) Evolution of the anomalous Hall resistance or the weight of the magnetic synapse under the application of consecutive gate-voltage pulse of +5 V magnitude and 50 ms duration. Various curves represent the device evolution for the different values of time delay between each pulse. Readings are taken after every 20 pulse-application events. (b) Dependence of the synaptic weight modulation on the relative timing of the pre and postsynaptic pulses. Solid lines represent an exponential fit.

Figure 4

(a) Learning and forgetting curve of the magnetic synapse. Gate voltage is set to 0 V while measuring the anomalous Hall resistance during the forgetting or relaxation process. (b) Forgetting or relaxation process after consecutive learning events that do not include any training step. Solid lines represent fits using Eq. (1).

Figure 5

(a)–(d) Forgetting or relaxation curve of the magnetic synapse after applying a different number (N) of training pulses. Training pulses are applied after setting the magnetic synapse to its maximum weight. Solid lines represent the fits using Eq. (1). (e) Relaxation time as a function of the number of training pulses.