In biological neural networks, neurons are connected to other neighboring neurons via synapses, as shown in Figure 1A. Neurons integrate input spike signals from adjacent neurons via synapses, i.e., integrate. In addition, neurons generate output spike signals when membrane potentials reach a threshold value, i.e., fire. This neuronal behavior is called “integrate-and-fire,” which is the key operation of neuron (Hodgkin and Huxley, 1952; Izhikevich, 2003)_. Similarly, artificial neurons could be connected with other artificial neurons via artificial synapses, where p-STT-based neurons are connected with memristor-type synapse, as shown in Figure 1B. The p-STT-based neurons receive spike signals through synapses connected with preneurons, integrate the signals, and then sends out output spike signals when the resistance of the p-STT-based neurons reaches a certain threshold value. In the following sections, we will describe in detail how p-STT-based neuron works.

Figure 2A shows schematic structure of p-STT MTJ. Its magnetic moment versus applied perpendicular magnetic field (M–H) loop was investigated to determine the static magnetic behavior of the p-STT MTJ, as shown in Figures 2B,C. It includes four perpendicular magnetic anisotropy (PMA) layers: a double Co₂Fe₆B₂ free layer (i in Figure 2A), Co₂Fe₆B₂ pinned layer (ii in Figure 2A), upper [Co/Pt]₃ SyAF layer (iii in Figure 2A), and lower [Co/Pt]₆ SyAF layer (iv in Figure 2A). Here, the Co₂Fe₆B₂ pinned layer was ferrocoupled with the upper SyAF layer, whereas the upper [Co/Pt]₃ SyAF layer was antiferro coupled with the lower [Co/Pt]₆ SyAF layer. The magnetic moments of the double Co₂Fe₆B₂ free layer, Co₂Fe₆B₂ pinned layer ferrocoupled with the upper [Co/Pt]₃ SyAF layer, and lower [Co/Pt]₆ SyAF layer were 0.130 (M_i in the inset of Figure 2C), 0.362 (M_{ii + iii} in Figure 2B), and 0.370 (M_iv in Figure 2B) memu, respectively. In addition, the double Co₂Fe₆B₂ free layer showed an excellent interface PMA characteristic with a good squareness and fair coercivity (H_c, ∼0.13 kOe), as shown in Figure 2C. This result indicates that the MgO tunneling barrier had good face-centered cubic crystallinity that enhanced the coherent tunneling of the spin electrons (Lee et al., 2016a,c,d). The magnetic resistance versus voltage (R–V) behavior at room temperature (295 K) was measured to investigate the spin transfer torque switching behavior of the p-MTJ, as shown in Figure 2D. The switching voltage from parallel to antiparallel states was −0.53 V (V_PtoAP), while the switching voltage from antiparallel to parallel states was + 0.61 V (V_APtoP). The magnetic resistance versus magnetic field (R–H) loop of the p-STT MTJ is shown in Figure 2E. When the applied perpendicular magnetic field was scanned from + 0.5 to −0.5 kOe, the electron spin direction of the double Co₂Fe₆B₂ free layer was rotated from upward to downward so that the electron spin directions between the double Co₂Fe₆B₂ free and pinned layers were switched from antiparallel to parallel states. As a result, the resistance of the p-STT MTJ decreased from 82 to 46 Ω. The squareness and coercivity of the p-STT MTJ measured with an R–H loop was almost the same as that measured with an M–H loop, indicating that this device could maintain a stable magnetic state in a zero magnetic field so that the integration behavior would be characterized during the switch from parallel to antiparallel between the double Co₂Fe₆B₂ free and pinned layers.

Interestingly, the p-STT MTJ showed integration property when consecutive voltage pulses (spike) were applied, as shown in Figure 3A. The spike width was 50 μs, and the spike amplitude was varied from −0.50 to −0.70 V. At all spike amplitudes, i.e., −0.50, −0.55, −0.60, −0.65, and −0.70 V, the p-STT MTJ performed the integration at input spikes of ∼100 pulses. In addition, the resistance difference increased when the input spike amplitude increased from −0.50 to −0.70 V at input spikes of ∼100 pulses, as shown in Figure 3A. Over an input spike amplitude of −0.7 V, no integration behavior was found. In addition, the p-STT MTJ showed a good repeatability for five sets of ∼100 input spike pulses, where the resistance increment by the 100 input spike pulses increased with the input spike, as shown in Figure 3B. Our proposed p-STT MTJ in Figure 3 showed a unique neuron characteristic (i.e., integration characteristic) compared to MTJ-based neurons (stochastic characteristic with a two-terminal device or leaky-integrate-and-fire characteristic with a three-terminal device), as shown in Supplementary Table 1. The mechanism of this behavior could be explained by understanding the grain-size distribution of the polycrystalline MgO tunneling barrier. The distribution of the sputtered polycrystalline MgO tunneling barrier was 0.6 to ∼1.8 nm, where the average grain size was ∼0.94 nm, as shown in Supplementary Figure 3. This indicates that even for a p-STT MTJ with a cell size of 35×35 nm², multiple grains would exist within the p-STT MTJ cell, as shown in Supplementary Figure 4. As a result, we can expect that the p-STT MTJ with a cell size of 35× 35 nm² would show an integrate characteristic similar to Figure 3 since it has a large number of grain within the p-STT MTJ cell. The interfacial PMA of both the double Co₂Fe₆B₂ free and pinned layers originated from the hybridization between O atoms and X (Fe or Co) atoms at the MgO tunneling barrier and Co₂Fe₆B₂ layer interface. Thus, the polygrain size distribution of the polycrystalline MgO tunneling barrier directly and strongly affects the ferromagnetic properties of both the double Co₂Fe₆B₂ free and pinned layers, i.e., resistance difference between parallel and antiparallel states of the p-MTJ. In addition, the hybridized Fe–O and Co–O bonds within the grains would be well oriented with the crystallinity of the MgO tunneling barrier, so the electron spins would require a high activation energy to switch from parallel to antiparallel. Otherwise, the spin electrons at the grain boundaries would have a relatively low energy barrier to switch from parallel to antiparallel, compared with the spin electrons within the grains (MacLaren and Willoughby, 2001; Victora et al., 2003; Kondo et al., 2018), as shown in Figure 4A. Thus, the spin electrons at the grain boundaries (Figure 4B) would first be switched from parallel to antiparallel states (Figure 4C), and the spin electrons inside the grain would then rotate due to the ferrocoupling between the spin electrons at the grain boundary and inside the grain (Figure 4D). As a result, the spin electrons in the grains would be switched from parallel to antiparallel, which would be a similar switching behavior to a previous report (Suzuki et al., 2016). This switching process would induce the integration behavior when the spikes are sequentially applied to p-STT MTJ (Figure 4E). The integration behavior of a p-STT MTJ was influenced by the crystallinity of the MgO tunneling barrier in Figure 2A, i.e., a better crystallinity of the MgO tunneling barrier led to a better integration characteristic, as shown in Supplementary Figure 5. This integration behavior of the p-STT MTJ would suggest that the p-STT MTJ could be applied with the complementary metal–oxide–semiconductor field-effect transistor (C-MOSFET) technology to produce artificial neuron. In general, the perpendicular spin torque switching time of a p-STT MTJ has been reported as ∼10 ns, which is the fastest switching time among other semiconductor devices (Hu G. et al., 2018). In addition, the operation of the integration by a p-STT MTJ in Figure 3 was performed prior to a full the perpendicular spin torque switching. Thus, the width of a spike pulse in Figure 3 could be less than ∼10 ns if the size of a neuron using a p-STT MTJ can be scaled down–up to 35 × 35 nm², suggesting a lowest power consumption per a spike in neuron (i.e., 1.6 × 1.6 μm²), as shown in Supplementary Table 2.

Although the p-STT MTJ exhibited integration behavior depending on the input spike amplitude, it requires an additional circuit to perform the fire operation. Thus, the p-STT MTJ-based neuron circuit was designed using one p-STT MTJ, seven n-MOS FETs, three p-MOS-FETs, and one reference resistance to conduct the integrate-and-fire operation as shown in Figure 5A. Note that we calculated the area of the p-STT MTJ-based integrate-and-fire neuron using 1.6×1.6μm² p-STT MTJ (i.e., ∼8.2 μm²), which was approximately one-fourth smaller than the previous report (Sourikopoulos et al., 2017), as shown in Supplementary Figure 6 and Supplementary Table 3. In this circuit, “fire” occurs when the resistance of the p-STT-based neuron exceeds the reference resistance (R_ref). The neuron receives control signals from a controller and performs integration, read, and reset operations in each clock cycle, as shown in Figure 5A. One controller can control multiple neurons simultaneously. In order to implement neural network, cross-point array can be used to realize analog matrix-vector multiplication. Figure 5B shows the schematic illustration of typical cross-point neural network implementation, which was fabricated by a cross-point synapse array being connected with our proposed p-STT MTJ neuron. Synapse would be IGZO-based memristor (in our experiment shown in Supplementary Figure 2). Where the bias voltage (V_bias) serves to ensure that the p-STT-based neuron is within its proper operating range.

To investigate the performance of the SNN, a single-layer SNN consisting of input and output layers (50 p-STT-based neuron) was designed, as shown in Figure 6A. In this simulation, IGZO-based ReRAM was used in artificial synapse. A performance test of the SNN was carried out using the MNIST handwritten image set. MNIST images (6 × 10⁴) were used for training, and 1 × 10⁴ images not included in the training were used for testing. The probability of the input spike occurrence was set to be proportional to the pixel value of an input image, and the amplitude of an input spike was set to −1 V. The neurons integrate the input spike signals and fire when the resistance of the p-STT MTJ exceed R_th (=70 Ω). When the neurons fire, they generated an output spike. The winner takes all (WTA) was applied to the output neuron nodes. WTA improved the accuracy of a single-layer SNN since the WTA guarantees non-linear mapping in a single-layer SNN (Du et al., 2015; Hansen et al., 2017). Finally, only the synaptic weights associated with the fired output neurons were updated. In the initial synaptic weight map, the conductance of the synapses was randomly distributed. After training, the distribution of synaptic weights was changed. The weights for active and silent neurons are shown in Figures 6C,D, respectively. Even if there were more than 10 epochs, there were some silent neurons, as shown in Figure 6D. These silent neurons exhibited almost no firing during training. The reason for this is that the WTA updates only synaptic weights associated with neurons that have fired; consequently, synaptic weights connected with neurons that rarely fire are slower to learn. As a result, these less learned synapses reduce the firing rate of the connected silent neurons compared to other neurons. In the end, learning is rarely achieved for the silent neurons. In biological neural networks, there is a mechanism called “homeostasis” to overcome these problems. With this mechanism, a neuron that frequently fires increases the threshold required to fire, and a neuron that rarely fires decreases it (Lee et al., 2016b,d; Johnson et al., 2018). This mechanism lowers the fire threshold of neurons where learning has not been achieved; thus, it causes neurons to be more likely to fire during subsequent learning. However, it is difficult to change the reference resistance R_th once it is set in the circuit. This remains a problem to be solved in the future. We use simplified STDP learning rule for synaptic learning. The synaptic weights before training are shown in Figure 6B. First, we simulated the dependence of pattern recognition accuracy on read error using our proposed the cross-point synapse array (i.e., Figure 5B) being connected with our proposed p-STT MTJ neuron (i.e., Figure 5A), as shown in Supplementary Figure 7. The pattern recognition accuracy sustained at ∼76% up to read error of 5% and then rapidly decreased with read error larger than 5%. In addition, we tested the dependence of pattern recognition accuracy on the reference resistance by simulation, as shown below Supplementary Figure 8. We determined the reference resistance that showed the highest accuracy of pattern recognition simulation. Using the simulated reference resistance, the pattern recognition accuracy rapidly increased to ∼76% in two epochs, as shown in Figure 6E. Since the single-layer SNN used in training is learned through STDP unsupervised learning, so only clustering was performed for each output stage. Therefore, the most frequent output values of each node were compared with the determined input value to measure the pattern recognition accuracy. The single-layer SNN, composed of p-STT-based neurons, showed a maximum recognition accuracy of ∼76%, which was somewhat lower than that of other reported neural networks (Burr et al., 2014). In the single-layer SNN, pattern recognition accuracy increases with the number of output neurons (Querlioz et al., 2015; Zahari et al., 2015; Hansen et al., 2017). However, even if the number of output neurons is increased to 100, it is difficult to obtain more than 90% accuracy. The major reason for the low accuracy is the lack of proper learning algorithms to train SNN. The spike signals are not differentiable, so global learning rule such as backpropagation cannot be used for training SNN. Therefore, local learning rule such as STDP is mainly used for training SNN. This limits the structure of neural network to a single layer. Therefore, in order to increase the accuracy of the SNN, further study of the learning algorithm is necessary.