Voltage-controlled interlayer coupling in perpendicularly magnetized magnetic tunnel junctions

Magnetic interlayer coupling is one of the central phenomena in spintronics. It has been predicted that the sign of interlayer coupling can be manipulated by electric fields, instead of electric currents, thereby offering a promising low energy magnetization switching mechanism. Here we present the experimental demonstration of voltage-controlled interlayer coupling in a new perpendicular magnetic tunnel junction system with a GdOx tunnel barrier, where a large perpendicular magnetic anisotropy and a sizable tunnelling magnetoresistance have been achieved at room temperature. Owing to the interfacial nature of the magnetism, the ability to move oxygen vacancies within the barrier, and a large proximity-induced magnetization of GdOx, both the magnitude and the sign of the interlayer coupling in these junctions can be directly controlled by voltage. These results pave a new path towards achieving energy-efficient magnetization switching by controlling interlayer coupling.

Black is after Vset > 0, setting the system in a FM state after the initial AFM state (Blue), Red after Vset < 0, resetting the system to the initial AFM state. The change in HIC state between AFM and FM is nonvolatile. We use elevated temperature and high bias voltage to change the state but 4 the state is then measured at RT and low bias. To test the retention of this effect we have measured the HIC as a function of time after the most recent state change. The HIC is maintained for at least 3 weeks (as long as the test was run for). Error bars are the standard deviation of 3 repeated curves at each data point. This retention test was carried out on a sample with 2.5 nm GdOx.

Supplementary Note 2: VCMA effect
In addition to the VCIC effect discussed in this paper this system is the first other than demonstrating that the change of HIC is indeed due to the voltage applied to the pMTJs.

Supplementary Note 4: Dependence of HIC with PMA of GdOx-pMTJ
According to a model proposed by Moritz et al. 6 , the interlayer coupling in (Co/Pt)-Ru-(Co/Pt) PMA spinvalves can be explained by an extended Neel-type orange peel coupling induced by correlated roughness. In this model, the AFM coupling is stronger in samples with larger PMA, which cannot explain the decrease of AFM coupling illustrated in Supplementary   Fig. 8. A pMTJ was measured after annealing at 260 °C for 60 s, then subsequently after 11 annealing at 300 °C for 180 s. The PMA of both of the magnetic layers was higher after the 300 °C annealing, as is evident from larger switching fields for both the hard and soft CoFeB layers as shown in Supplementary Fig. 8a. This increase in PMA, however, corresponds to a decrease in AFM coupling as shown in Supplementary Fig. 8b, with HIC dropping from -40 Oe to -31.5 Oe after 300 °C annealing. This suggests that the AFM coupling in the CoFeB/GdOx/ CoFeB system cannot be explained by the Moritz model. (1)

Supplementary
The explicit forms at M4, 5 edges for 4f rare earth metals are: where μ+ (μ-) is the absorption intensity with left (right) circularly polarized x-rays; μ0 was Due the much larger Fe concentration in the Co20Fe60B20 electrodes, the magnetic field dependence of XMCD signal was only performed for Fe. 13 In order to further confirm that the large Gd magnetization is due to the proximity effect, we have measured the dichroic reflectivity of two unpatterned reference samples: a CoFeB/GdOx bilayer and a GdOx single layer. The resonant magentic scattering data has much better signal/noise ratio making it very sensitive to the weak magnetic signals from buried interfaces 12,13 .Experiments were conducted with an incident angle of 10 deg. with ±5 kOe inplane fields. The data shown in Supplementary Fig. 10 were normalized by the direct beam intensity. There is a strong magnetic scattering near the Gd M5 edge in the CoFeB/GdOx bilayer, but not in the single GdOx layer sample, indicating that the strong Gd magnetic signal is indeed due to the proxmity effect.

Supplementary Note 6: Possible origin of the observed VCIC effect
Previous theories [14][15][16][17] of voltage controlled interlayer coupling did not consider the direct manipulation of HA and MS or, more significantly, a large induced magnetic moment in the barrier. Therefore, these theories cannot be applied to the observed effects in the present GdOx pMTJs. The average ferromagnetic moment of Gd ions induced by CoFeB is 0.6 μB per Gd 3+ ion, which is nearly 20 times larger than the induced moment of Pt in Pt/Fe bilayers 18 . Interfaceinduced magnetization in tunnel barriers has also been observed in other MTJ systems, where it affects the spin-dependent transport 19 . We do not expect, however, the induced moment of Gd 3+ to significantly affect the transport properties of our junctions due to the localized nature of f electrons. It may, however, play an important role in the magnetic properties, e.g., the coupling of the two CoFeB layers in the pMTJ. Generally in the magnetic proximity effect, the depth distribution of the induced moments in the NM layers varies from system to system. For example, it is estimated that 90% of the induced moments in Pt on Co exist only in the first four monolayers from the interface, with a characteristic decay length of 0.41 nm 20 . In other cases, however, the induced magnetization can extend much deeper into the NM layer, such as in Bi2Se3/EuS where the first 2 nm of Bi2Se3 was found to be ferromagnetic 21 .
The thickness dependence of the HIC in the initial state of the GdOx-pMTJ is plotted in Supplementary Fig. 12. The AFM coupling in our samples is extended to much thicker barriers Under present conditions, VSET (0.5-0.8 V) needs to be applied to the GdOx-pMTJs for between a few tens of seconds and a few minutes, which could lead to the breakdown of the tunnel barrier when the GdOx is less than 2 nm thick. Due to this limitation, most VCIC experiments in this work were performed on pMTJs with GdOx thicker than 2 nm. This situation in principle can be improved once GdOx with higher quality can be fabricated. Therefore we expect to observe the same VCIC in junctions with thin barriers (GdOx < 1.5 nm) when the quality of the GdOx barrier is improved.
In the following, we describe the VCIC with thick GdOx (> 2 nm) with a model considering the voltage-driven oxidation level changes of Fe, the large induced moment of the 15 Gd ions that is proportional to the amount of free Fe, and a voltage dependent distribution of correlated moments in the Gd ions. Since we have shown that oxygen in the GdOx barrier can be reversibly moved toward or away from the interface by applying voltage, we propose that the correlation of the magnetic moment distributions between the two interfaces created by the oxygen transportation may be responsible for the observed VCIC. The XMCD has shown that the Gd ions display significant magnetic moments due to the proximity effect with CoFeB. These large induced Gd moments may contribute to the interplay coupling via dipolar interaction, especially in samples with thick barriers (> 2 nm).
Consider two thin magnetic layers separated by a distance d. The coupling energy between these two layers can be expressed as: where 1 ( 2 ) is the magnetic moment at the position 1 ( 2 ), 12 = 1 − 2 +, 12 = | 12 |, and ̂1 2 = 12 / 12 . For perpendicularly magnetized layers, the above integration is identically zero if the magnetic moments in each layer are uniformly distributed. In the present case, the migration of the oxygen vacancies to the two interfaces may create a highly non-uniform distribution of the interface moments. Let's introduce an impurity moment ( ) = ( ) − ̅ , where ̅ is the optimally and uniformly magnetized magnetic moment of the i-th layer. One may think of as a Gd ion that is polarized by a Fe atom next to it. Consequently is zero if the neighboring Fe atom is oxidized by O 2driven by the applied voltage. To estimate the magnitude of the coupling from such non-uniform distributions, we define the correlation function between the distributions of the magnetic moments of the two layers, 12 ( 1 − 2 ) =< 1 ( 1 ) ⋅ 2 ( 2 ) > −< 1 ( 1 ) >⋅< 2 ( 2 ) > (6) where <> refers to the spatial average over the plane of the layers (for a fixed 1 − 2 ). Due to the much stronger PMA of the bottom CoFeB layer (Supplementary Figure 3), the change in magnetic properties of top CoFeB layer by voltage to be much larger than that of bottom CoFeB, which is supported by Fig. 4b. We assume that the applied voltage from the positive to negative polarity results in a change of the correlation function from the correlated ( 12 > 0) to the anticorrelated ( 12 < 0) state, i.e., the oxidation level (therefore the induced moment from Gd ions) of the two interfaces is likely similar if | 1 − 2 | < 0 for a positive voltage, where 0 is the correlation length. For a negative voltage, the oxidation level of the two interfaces is likely to be less similar, giving rise to 12 If we take a simple correlation function 12 ( ) = 2 2 ( 0 − ) where n is the density of impurities, S is the spin, and is the step function, one can analytically integrate the above equation. In the inset of Supplementary Fig. 13, we show the coupling field as a function of the correlation range 0 . Taking the maximum value for the optimal 0 for each thickness of the barrier, we show in Supplementary Fig. 13 the coupling field as a function of the barrier thickness with = 1 and = 3.

17
The GdOx barrier in the pMTJ used in the XMCD experiment has a thickness of 3.4 nm, with an average induced moment of 0.6 μB per Gd ion as shown in Supplementary Note 5. If most of these moments reside in only the first monolayers of GdOx (0.27 nm) next to CoFeB on both sides of the barrier, the induced moment per Gd ion can be as large as ~3.7 μB, which is the basis for using S = 3 in Equation 7. Alternatively, if we assume that the induced moment of the Gd ions resides in the first two monolayers (0.54 nm) next to CoFeB on both sides of the barrier, we have an average induced moment of ~1.8 μB per Gd ion, which can be described approximately by S = 2. For = 1, S = 3, and = 2 nm, we find HIC to be about 2500 Oe as shown in Supplementary Fig. 13. With a more realistic approximation of n = 0.2 that corresponds to a 20% impurity density, we have HIC ≈ 100 Oe with the sign determined by ξ12, which is comparable the experimental results.
In conclusion, while the simple model based on the voltage driven correlation of the magnetic moments of two magnetic layers can roughly account for the sign change and the magnitude of HIC, there are parameters that are not fully justified. In particular, the detailed correlation function is unknown at the present time. Further experiments are needed to establish definitive mechanisms for the unique VCIC observed in GdOx-pMTJs.