For MTJ, a field sweep test and current-induced switching under an in-plane field (SOT-switching) are performed to obtain fundamental MTJ properties. Then, pulse application under a zero field is done to show DW movement across the MTJ. Finally, sequential positive and negative pulse application alternately is implemented to prove reliable DW motion in an IBE damage-cured device.
Before we dive deeper into our main results, we briefly verify the advantage of DW-MTJ in yield and writing energy point. Figure 1a illustrates the short ratio of the device as a function of etching time, measured from the MgO peak in the signal of an optical emission spectrometer, with schematics of each device. We measure parallel-state resistance R and count the number of MTJs that have R less than half of the normal device. The short ratio is then obtained by dividing the number of measured devices (48). The result clearly indicates that the yield of DW-MTJ is higher than SOT device.
Figure 1b, c show the current-induced SOT switching and DW propagation investigated by MOKE microscopy, respectively. For SOT switching, the device is first uniformly magnetized by a large enough perpendicular field H and then one pulse with 20 ns duration is applied under a magnetic field of current direction H with magnitude 30 mT. For DW-propagation, DW is initially installed on both edges of the device, and 10-shot pulses with 20 ns under zero field are applied. The displacement region is investigated as the area where the contrast changes. The DW propagation is observed from much lower current density than that where SOT switching occurs, indicating that our experimental results are consistent with what we stated in the introduction.
Figure 2a shows the perpendicular field H dependence of MOKE signals in the half-stack. The signals displayed in the figure are obtained from different CoFeB thicknesses. CoFeB and Co/Ni multilayers are coupled antiparallel, resulting in virtually zero net magnetic moment. Figure 2b presents typical MOKE images of a DW moving along the device. Before starting pulsation, DW is installed at the edge of the line region as indicated by the red broken line on the left, and then that state is recorded as the initial state for differential image acquisition. The current pulse with 2 ns duration is applied 50 times, and a differential image from the initial state is saved. Figure 2b shows the saved images in the case of current densities 34.7, 52.0, and 104 MA/cm. Darker contrast between the red broken lines indicates DW displacement. We evaluate the displacement length by detecting the abrupt change in grayscale when viewed along the longitudinal direction within the line. From the obtained displacement length, we calculate v as displacement length/(pulse duration × shot number). Figure 2c summarizes v in many devices having different CoFeB thickness. v dramatically depends on CoFeB thickness. The cause of the trend is not fully understood. We believe that it is due to two factors occurring simultaneously: one is the contribution of magnetization compensation, and the other is the dependence of DW speed on anisotropy. v is lower when the anisotropy of the layer is larger, and in the case of CoFeB, perpendicular anisotropy increases with decreasing its thickness due to its interfacial anisotropy. Therefore, v reaches its highest value when the CoFeB is slightly thicker than the compensation point, and as the thickness and hence the anisotropy increase, the speed decreases. By tuning the stacks, v can reach about 200 m/s with a current density J less than 100 MA/cm, indicating that Co/Ni multilayer and CoFeB/MgO SAF structure have high potential as DW-MTJ materials. Using the obtained v and J values of 200 m/s and 60 MA/cm, we estimate the necessary pulse duration for write operation t and current I as a function of device length L and width W, as displayed in Fig. 2d, e. From the half-stack experiment, it is obtained that I ~ 252 μA at a pulse width of t = 0.5 ns when the device size shrinks to a length of 100 nm and a width of 50 nm. The obtained values in this study are lower than the previous study reported fast DW speed ~750 m/s in SAF structure; however, as shown in this calculation, DW speed of 200 m/s is already fast enough to approach SRAM replacement.
DW-MTJs utilize DW propagation from an existing domain. For the switching operations, the device should be designed so that spins aligned in opposite directions exist at the two ends of the element. Since a global field cannot create this antiparallel alignment, this initialization process is an important challenge for DW-MTJ. A previous study suggested adding pinning layers underneath FL to control magnetization at the two ends independently, but it is difficult to make the magnetization directions of these layers antiparallel in real applications.
We propose a new reliable technique, the concept of which is illustrated in Fig. 3a, b. The initialization is done by applying multiple current pulses of the same amplitude to the device under an in-plane magnetic field H. The first pulse, with a sufficiently large amplitude, induces SOT-switching; magnetization in the narrower linear region flips, and two DWs are installed between the linear region and the triangular parts (Fig. 3a). Subsequent pulses push the DWs right by current-induced DW motion, as shown in the dashed arrows in Fig. 3b. In addition to this, however, spin currents by SOT always uniformly switch the linear region. Therefore, one DW stays at the edge of the linear region (the left DW in Fig. 3b), while the other moves far into the triangular region (the right DW). Figure 3c-g demonstrates this scheme. As explained, the left-side DW does not move after 10 shots, while the right DW travels to the rectangular region. Since the current density of the triangular region is much smaller than that in the linear region, the installed antiparallel spin configuration in the two triangular regions is not affected by normal operation.
Note that this initialization protocol is only performed when the device is fabricated using a global magnetic field, and the individual customer only performs DW motion back and forth for writing. The field applied during the experiment, in fact, is not necessary in a real device. Thus, this device scheme is what we call the scalable design.
Figure 4 summarizes the stability of this initialization method. We installed two DWs utilizing the procedure we described above, and then a 20-shot pulse with 60 MA/cm was applied to move the left DW to a suitable position for the following measurement. As depicted in Fig. 4a, we call the left DW as DW1 and the right one as DW2. A reference image is taken, and the pulses that drive DWs from right to left are applied. The pulse duration is 20 ns, 10 shots, and the applied field is zero. The results with varying the amplitude are shown in Fig. 4b-g. As can be seen, DW1 starts to move from 36 MA/cm, but DW2 starts to move from -119 MA/cm. Since DW2 should not move for reliable operation, the write pulse should not exceed 119 MA/cm. Considering that normal operation is done at maximum DW speed with the current density of 72 MA/cm, the operation margin defined as the gap between operation current density and DW2 moving current density is large enough. In order to check the scalability of this method in a small device, we perform numerical calculations. Figure 4h shows the distribution of current density along the current direction versus the ratio of the width of the triangle part to the width of the linear region. DW2 locates 5.5 mm far from the edge, and the current density becomes ~28% of that in the middle of the line. It denotes that DW2 is applied ~ -20 MA/cm when DW1 is assigned -72 MA/cm where the highest speed is obtained. Figure 4i summarizes the obtained current density dependency of DW1's speed and illustrates the current densities applied to DW 1 and DW2 in the states of (c) and (e). As depicted, the current density applied to DW2 is low enough that DW2 remains in the triangle region and is consistent with experimental results. In this paper, we do not show, but further improvement of the design can be considered, such as shortening the triangle region to vanish DW2 or making a stable pinning site in the region.
Figure 5a shows MOKE results for a sample before and after IBE, and after an additional anneal for curing IBE damage. The anneal cure is performed at 300 C for 1 h under vacuum. Comparison of these results suggests that the SAF magnetization structure is destroyed due to IBE damage, and the after-anneal result validates that additional anneal recovers perpendicular magnetic anisotropy (PMA). According to a previous study, ion etching introduces the ion to MgO and/or MgO/CoFeB interface and induces degradation of magnetic properties. In the property recovery investigated here, we expect that those introduced ions are removed during partially activated solid-state epitaxial growth between MgO and CoFeB. Figure 5b-g display MOKE images of DW motion in the device without (b-d) and with anneal (e-g). In the device structure shown in Fig. 5b-d, the MTJ cells block the ion beam during the IBE process, and the region underneath the cells is not damaged by the ion beam. However, the region that is not covered by the MTJ cells is affected by the ion beam. Especially, the top layer of SAF is damaged and magnetization turns into nearly in-plane, breaking the SAF property, as shown in Fig. 5a after the IBE case. The resulting track line is not magnetically uniform but a type of lateral junction composed of single (non-covered region) and SAF (MTJ cell covered region) layers. In this non-uniform structure, DW can be nucleated at the boundary between these two different magnetic structures, as one can see in the DW motion in Fig. 5b-d. After the recovery of magnetic properties by anneal cure, on the other hand, the track line becomes magnetically uniform and keeps SAF behavior all along the line. The DW moves smoothly without disturbance from the existence of MTJ cells, as it is shown in Fig. 5e-g. This gives clear proof that IBE damage strongly affects the DW motion, and anneal recovers it. Since MTJ cells are an essential part of the device reading, the anneal cure process is very important for the device.
Utilizing the initialization technique and damage cure method discussed above, we demonstrate repetitive write operations in DW-MTJ devices. Cross-sectional transmission electron microscopy image and corresponding schematic illustration of our device are presented in Fig. 6a, b. Figure 6c shows the H dependence of TMR, and Fig. 6d depicts the results of current-induced magnetization switching detected using TMR in DW-MTJ. The CoFeB/ MgO structure is chosen to obtain a high TMR of about 150% as reported; however, in our devices, TMR values varied between 50-100%. The discrepancy will be improved by the polishing process technique and/or a slight modification of the stack. Here, we apply pulses of 200 ns duration under μH = 50 mT. Applied H is larger than the previous experiment above, but this is simply because it makes SOT switching more reliable. Probably, if it is reduced to the value of a half-stack device it only requires more current for switching. The pulse amplitude is denoted in current density J calculated from commanded voltage, device resistance, device width and thickness. The switching measurement is done 10 times, and all results are shown in the figure with different colors. To show fundamental switching properties, the switching probability distribution of 100 k trials as a function of pulse width t and t dependence of switching current density J are summarized in Fig. 6e, f, respectively. Note that the data in Fig. 6e, f are obtained from a different device compared to Fig. 6c, d. As can be seen from the figure, the deviation of switching current density is small and stable switching can be expected in long enough t.
The write operation characteristic shown in Fig. 7a is obtained as a function of applied pulse number and pulse amplitude. The measurement implemented here is done after initialization with application of 75.4 MA/cm, 200 ns pulse under μH of 50 mT, which we proposed in the "Initialization technique" section above. The write pulse is fixed at 5 ns. As can be seen from the figure, the TMR value changes after applying several pulses of appropriate amplitude. This demonstration denotes that our proposed initialization technique and damage cure method are working properly at the DW-MTJ level. From this result, v can be estimated as 73.2 m/s at commanded voltage V = 0.45 V using the distance from DW initialization position to MTJ edge 366 nm. Corresponding current and current density are calculated as I = 2.26 mA and J = 67.9 MA/cm at V = 0.45 using separately measured current path resistance R = 349 Ω. Finally, we demonstrate repetitive write operations after initialization of one DW utilizing the method discussed above. Figure 7b, c shows the results obtained from a device with a length of 500 nm. An applied pulse duration of 10 ns drives the DW to the other end. Most of the time, the write operation succeeds; however, Fig. 7c indicates it sometimes fails to write, suggesting further quality improvement is necessary. Note that we perform the same experiment on the other DW-MTJ having the length of 2000 nm and the width of 900 nm, and obtain DW moves across MTJ in one shot of the pulse with an amplitude of 56 MA/cm and a pulse duration of 20 ns. It is found that there is a roughly linear relationship between device length and the necessary pulse duration that drives DW to the other end in one shot. According to this relationship, SRAM-level operation speed, which is typically 1 ns can be achievable by reducing device length down to 100 nm under the assumption that the fabrication quality can be maintained at a similar level. Moreover, DW experiments in DW-MTJ shown in this study are all one-time measurements despite of SOT switching experiment, thus we cannot discuss statistical variation. Discussing the statistical distribution of properties is highly required for industrial applications, and we believe that further research on these topics is necessary in this field. On the other hand, we state that our device structure is more forgiving to pinning, which is assumed to be installed during nm-size device fabrication. DW always moves back and forth in the linear region and consistently stops at the boundary between the linear and triangular region, illustrated in Fig. 6b, due to the abrupt change of current density caused by the sharp change in width. The sharp transition ensures that the DW cycle is highly stable against pulse amplitude variations. Pinning sites arising from device fabrication processes are expected; however, their effect is anticipated to be smaller compared to the sharp width transition in the device. Therefore, by increasing pulse amplitude until pinning sites caused by etching-induced inhomogeneities become negligible, the error can be effectively suppressed. Additionally, we comment to DW retention for future application. According to a previous study, thermal stability factor E/kT had been demonstrated in the devices with sizes below 50 nm, which was comparable to or even better than that of spin-transfer torque (STT) devices. Although we cannot ensure that the same value will be achieved since we use a different material system, we expect that the current technology can satisfy the retention requirements for devices with sizes on the order of several tens of nanometers.
Experimental results in this and previous studies are compared in Table 1. Although we adopt the SAF structure as a free layer with CoFeB for high-speed DW motion and high TMR, we acknowledge that neither DW speed nor TMR ratio reaches reported values, probably due to the lack of established processing techniques for this device. However, the significance of our findings is that the proposed structure is that the device does not require a pinning layer of DW initialization nor additional MTJ for STT write. The operation window is large enough to avoid interference from the other DW. The cure method properly recovers the magnetic properties of FL to be capable of operation. Here, we again mention that the proposed initialization is only once performed under the global field when the device is fabricated, and since the device does not contain an initialization scheme, DW cannot be re-nucleated after it accidentally annihilates. Therefore, the potential annihilation of the DW directory is a critical factor in determining device reliability, and thus, developing a stack that can sustain DW in the track for a sufficiently long period is highly necessary. We expect adding notches or adding a doping area in the triangular region can prevent annihilation of DW during usage, and anticipate that future research and development will enable us to meet at least the minimum requirements. In this study, we fabricated devices with sizes of 500 nm and 2000 nm, which exhibited operating pulse widths of 5 ns and 20 ns, respectively. By extrapolating the linear relationship between device length and operating pulse width, we expect that scaling down the device length to 100 nm will enable operation speeds of approximately 1 ns, rivaling those of SRAM.