Ferrimagnetic Heusler tunnel junctions with fast spin-transfer torque switching enabled by low magnetization

Main

Recently, magnetic random-access memory (MRAM), a non-volatile digital memory that utilizes magnetic tunnel junctions (MTJs) as the memory elements, was introduced as a foundry offering at the 28 nm node utilizing 100 ns write pulses¹. The MTJ is formed from CoFeB ferromagnetic electrodes that are separated by an ultra-thin crystalline MgO(100) tunnel barrier and exhibit very large tunnelling magnetoresistance at ambient temperature^2,3. The CoFeB is sufficiently thin that the electrodes exhibit perpendicular magnetization. One of the CoFeB layers acts as the memory layer and is written to ‘0’ and ‘1’ states (that is, opposite magnetizations) by the phenomenon of spin-transfer torque (STT) when a large enough current is passed through the junction^4,5,6,7. The speed, however, is limited by the use of the relatively high magnetization of CoFeB owing to spin-transfer angular momentum conservation^8,9,10. Reducing substantially the magnetization of the memory layer would enable lower switching currents at higher speeds as well as a smaller access transistor, thereby enabling higher densities.

Several long-standing challenges have prevented the adoption of low magnetization materials. While many ferromagnetic materials exhibit low magnetization, most of them have low Curie temperatures (<400 K)^11,12. One approach to lowering the magnetization is the use of ferrimagnetic materials^13,14. Perhaps the best known ferrimagnetic materials are those composed of rare-earth elements (Gd, Tb) and transition metals (Co, Fe). However, these materials have low Curie temperatures and, moreover, a strongly temperature-dependent magnetization owing to the rare-earth magnetic component¹⁵, making them unsuitable for MRAM applications^16,17. By contrast, the ferrimagnetic binary Heusler compounds Mn₃Ga¹⁸ and Mn₃Ge¹⁹ have two antiferromagnetically aligned magnetic sub-lattices, each formed from Mn²⁰, a transition metal, that have a weakly temperature-dependent magnetization. Moreover, these compounds are tetragonal and, therefore, exhibit a strong volume derived perpendicular magnetic anisotropy. These properties make tetragonal ferrimagnetic Heusler alloys ideal candidates for high-speed magnetic memory applications. The challenges with these compounds are their growth as ultra-thin layers, with bulk-like properties, on technologically relevant substrates. Another challenge is to incorporate them as memory layers in MTJs that exhibit high TMR at a low resistance-area product (RA) suitable for STT switching²¹. In such ferrimagnets, the spin polarization of the tunnelling current may be compromised owing to the compensating nature of the two antiferromagnetically coupled Mn sub-lattices²².

Here we address these challenges and demonstrate STT-switchable MTJs with a memory layer formed from the tetragonal Heusler ferrimagnet Mn₃Ge^23,24,25,26. We demonstrate that these MTJs, which exhibit high TMR at low RA values ~10 Ωµm², have much lower switching current densities (~10 MA cm⁻²) at short write pulse lengths of 0.5 ns, than for MTJs formed from conventional ferromagnetic electrodes for otherwise comparable thermal stabilities^27,28. Moreover, these Mn₃Ge-based MTJs are formed on amorphous SiO_x and are thermally stable at temperatures above 400 °C, making them compatible with typical complementary metal–oxide–semiconductor (CMOS) back-end-of-line processing.

Ordered alloy growth on amorphous silicon substrates

A recently developed chemical templating layer (CTL) technique^29,30 was used to grow the Mn₃Ge-based MTJ stacks, but here we refine the technique to allow for growth on Si/SiO_x substrates, which is required for CMOS compatibility. CTLs are binary compounds with the CsCl structure having a (001) texture to promote chemical ordering in single-unit cell thick Heusler alloy films deposited on them, even at ambient temperatures²⁹. They also enable the desired perpendicular magnetic anisotropy (PMA) in the Heusler layer²⁹. This previous work²⁹ used single-crystalline MgO(001) substrates so that such structures are not readily compatible with CMOS technologies. Promoting the growth of the CTL on amorphous SiO₂ layers is not straightforward. After conducting an extensive exploratory search, we found that binary nitrides with a wide range of lattice constants and having an NaCl structure grow with a (001) texture on amorphous surfaces (Supplementary Table 1). Two of these nitrides, MnN (Supplementary Note 1 and Supplementary Figs. 1–3) and ScN, were selected for developing MTJ stacks. We found that metallic MnN was not as thermally stable as semiconducting ScN at ~400 °C, the annealing temperature required for CMOS back-end-of-line integration.

Cross-sectional high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) images, which are presented in Fig. 1a,b, show the epitaxial growth of a Mn₃Ge layer for an MTJ stack grown on Si/SiO₂ (stack details in figure caption). The CTL was grown over an ultra-thin ScN seed layer, here just ~1 Å thick. The CTL is formed from a bilayer of IrAl and CoAl, both of which exhibit a CsCl structure with a (001) texture, thereby templating the growth of a chemically ordered Mn₃Ge (001) layer on top (Fig. 1b). A CTL from a single layer of CoAl also is effective, but we found that higher TMR was possible by using the bilayer CTL. Similarly, ScN can be replaced by other nitrides (Supplementary Table 1). To further demonstrate the role of ScN, in Fig. 1c, we present X-ray diffractograms (XRDs) of the stack with and without any CoFeB reference layer, and with and without the ScN layer (see figure caption for the detailed stack descriptions). Note that a thick 400 Å Cr layer is included within the MTJ stack to allow for current in-plane tunnelling (CIPT) measurements, but the growth of the films is similar without this layer. With the ScN layer, well-defined Cr (002) and both CoAl and IrAl (001) and (002) peaks are observed, whereas without ScN, no such peaks are found, and instead both the CTL and Cr show a (110) texture, which cannot promote PMA in the Mn₃Ge layer. Thus, the CTL has the desired (001) texture only when grown on Cr/nitride or just nitride underlayers.

Ferrimagnetic Heusler tunnel junctions with fast spin-transfer torque switching enabled by low magnetization — **Fig. 1: Growth of crystalline CTL on silicon substrates.**

Tunnelling magnetoresistance measurements, DFT predictions and future outlook

MTJs grown in this way show Mn₃Ge free layers (FLs) with the desired magnetic properties, as seen in the resistance (measured by CIPT) versus out-of-plane (OOP) field (R–H) hysteresis loop in Fig. 1d. CIPT-TMR measured in this MTJ stack is 69% but when patterned into devices gives rise to higher TMR. The largest TMR that we find for STT-switchable Mn₃Ge layers is +87% (Supplementary Fig. 4). Note that the TMR is positive, that is, R_AP > R_P, which is opposite to previously reported results for MTJs with much thicker (300 Å) Mn₃Ge layers grown without a CTL²². Similarly, for thick (50 Å) Mn₃Ge layers grown with a CTL, we find a negative TMR with a higher value ~−109% (Supplementary Fig. 5). These thick Mn₃Ge layers are not current switchable owing to their very high magnetic anisotropy. The sign difference in TMR for thin versus thick Mn₃Ge layers reflects strained versus relaxed Mn₃Ge layers with the larger in-plane lattice constant and reduced tetragonality of the Mn₃Ge layer grown with the CTL. Indeed, the TEM images (Fig. 1a,b) show that the thin Mn₃Ge layer assumes the in-plane lattice parameter (~4.03 Å) of the CoAl CTL.

A positive TMR was found from density functional theory (DFT) calculations (Supplementary Fig. 6) for MTJs that include a Mn₃Ge layer with the same in-plane lattice constant as that found here for the CoAl CTL. For Mn₃Ge layers with the bulk lattice constant, previous DFT calculations show a negative TMR²². Note that the TMR sign is determined by which of the two Mn–Mn or Mn–Ge ferromagnetic layers have the larger magnetic moment, which, in turn, depends on the in-plane tensile strain according to DFT calculations. Furthermore, the spin polarization of the tunnelling current depends sensitively on this strain so that, surprisingly, the spin polarization from each of the Mn layers can have the same sign and, therefore, higher TMR. Indeed, our DFT calculations (Supplementary Note 2) show the possibility of very high TMR values of up to +400% (Supplementary Fig. 7).

We anticipate that higher TMR values are possible experimentally by further improvements in the structural perfection of the Mn₃Ge layer itself, as well as the interfaces between this layer and the MgO tunnel barrier. We have found that even the smallest degree of oxidation of the Mn₃Ge decreases considerably the TMR. This is ameliorated in the radio frequency sputter deposition of the MgO tunnel barrier here by deposition at large offset angles (>~30°) from normal incidence. Further improvements in the degree of chemical ordering of Mn₃Ge may be accomplished by alternative CTLs or using surfactants. The lattice mismatch between MgO and Mn₃Ge may be detrimental to TMR and may be improved by dopant engineering of Mn₃Ge to tune its lattice constant. Another avenue for improvement is via the use of large-scale, multi-chamber dedicated deposition systems, which helps prevent cross-contamination, and cryogenic deposition facilities, the use of both of which often leads to substantial (for CoFeB/MgO-based MTJs) improvements in TMR as compared with laboratory-scale exploratory deposition systems, as used here.

STT switching, magnetic properties and energy barrier of Mn₃Ge FL

STT measurements were performed on circularly shaped devices that are 30–40 nm in diameter. A typical ~36 nm diameter device is shown in the cross-sectional image presented in Fig. 2a. The Mn₃Ge forms the FL. The reference layer (RL) is formed from CoFeB that is ferromagnetically exchange-coupled, via a thin Ta spacer layer, to a synthetic antiferromagnetic (SAF) structure composed of two distinct (Co/Pt) multilayers separated by a thin Ir layer^31,32. The two memory states of the MTJ correspond to the moment of the Mn₃Ge FL being parallel (P) or antiparallel (AP) to the moment of the lower layer of the RL, with corresponding resistance values, R_P and R_AP, respectively.

**Fig. 2: Stack details and switching characteristics of a Mn₃Ge-FL MTJ device.**

Figure 2b,c shows the switching between the P and AP states of the MTJ, driven by field (b) and current (c). The MTJ is switched reversibly by using an OOP magnetic field (Fig. 2b), resulting in two well-defined, non-volatile states in the absence of an external magnetic field. By initializing the net moment of the SAF along +z, we infer from the magnetic hysteresis loop that the higher resistance belongs to the AP configuration. Therefore, the resulting TMR ((frac{{R}_{{rm{AP}}}-{R}_{{rm{P}}}}{{R}_{{rm{P}}}})) is positive, exhibiting a value of +45%. In Fig. 2c, the STT-driven switching of this device is demonstrated by applying voltage pulses of increasing amplitude. Reversible current induced switching back and forth between the P and AP configurations is clearly observed. From the resistance versus voltage (R–V) hysteresis loop, it can be confirmed that the Mn₃Ge FL is switching.

The magnetic properties of Mn₃Ge thin films were studied by growing stacks without the RL. The magnetic OOP hysteresis loops for a series of Mn₃Ge thin films with varying thickness ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}) are shown in Fig. 3a. All the films exhibit PMA with high remanence. We find that the saturation magnetization (M_s) and coercivity (H_c) are very sensitive to small changes in ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}) (Fig. 3b). H_c more than doubles for every 2 Å increase in thickness reaching a very high value of ~42 kOe for ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}=21{mathring{rm{A}}}). Being able to attain such high values of H_c can be beneficial for making devices impervious to external fields. The H_c for thicker films exceeds 70 kOe, the maximum field in our measurement apparatus. For ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}=11{mathring{rm{A}}}), M_s is ~165 emu cc⁻¹, close to the bulk value²⁶, but decreases as ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}) is further increased. This observation likely arises from a subtle shift in the delicate balance of magnetic moments in the Mn–Ge and Mn–Mn layers. The H_k values were extracted from in-plane magnetic hysteresis loops (Fig. 3d). For ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}} > 17{mathring{rm{A}}}), H_k was too large (>70 kOe) to be measured (Supplementary Note 3).

**Fig. 3: Magnetic measurements and estimation of E_B for various thicknesses of Mn₃Ge (({boldsymbol{t}}_{{mathbf{Mn}}_{mathbf{3}}{mathbf{Ge}}})).**

A useful parameter is the thermal stability factor that is given by (Delta =frac{{E}_{mathrm{B}}}{{kT}}), where E_B is the energy barrier to switch the magnetic FL volume (V_m) between the P and AP states and k is the Boltzmann constant. For MRAM applications, ∆ should exceed ~50 to meet data retention requirements. For single magnetic domain reversal (macrospin approximation), (Delta =frac{{M}_{{rm{s}}}{V}_{{rm{m}}}{H}_{{rm{k}}}}{2{kT}}), but, typically, smaller values are found, except in small devices (<30 nm)^33,34. Experimentally, ∆ can be extracted from STT-driven MTJ switching experiments when the switching voltage V_sw is less than V_C0, the threshold voltage for switching in the absence of thermal fluctuations^35,36. V_sw is extracted from R–V scans (see Fig. 2c for the case of 0.5 ms long pulses). V_sw is plotted for voltage pulse length (t_p) ranging from 0.5 ms to 100 ms in Fig. 3e for a 30 nm diameter device with ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}=17{mathring{rm{A}}}). Within the macrospin approximation^35,36,37, the slope of V_sw versus (mathrm{ln}left(frac{{t}_{mathrm{p}}}{{t}_{0}}right)) gives (frac{1}{Delta }). Values of ∆ thus obtained for both P → AP and AP → P switching processes are then averaged, and a value of ~60 is estimated for this device. The mean ∆ for devices with a diameter of 35 nm is plotted versus ({t}_{{{rm{Mn}}}_{3}{rm{Ge}}}) in Fig. 3f. These values are in good agreement with estimates based on the values of M_s and H_k deduced from Fig. 3b,d. Our results show that Mn₃Ge FLs may enable scaling the diameter of such MTJs to below ~10 nm.

Reliable high-speed switching

Having established that Mn₃Ge FLs have high thermal stability from millisecond STT-switching characteristics, we now discuss their current-driven switching properties at much shorter timescales (≤10 ns), where thermal fluctuations play little role. Instead, the reversal takes place above the threshold current (({I}_{{rm{C}}0}=frac{{V}_{mathrm{C0}}}{{R}_{mathrm{p}}})) at which the STT surpasses the intrinsic damping torque^38,39. In this precessional regime, the switching becomes faster with increasing current. At a given temperature, the reversal starts from a thermally distributed initial state. The switching current I_c for a given write error rate (WER) is given by the following equation^9,37,39:

$${I}_{mathrm{C}}=frac{4e}{{mu }_{mathrm{B}}{gP}}left(alpha gamma {E}_{mathrm{B}}+frac{{M}_{mathrm{s}}{V}_{{rm{m}}}}{4{t}_{mathrm{p}}}log left(frac{{pi }^{2}Delta }{4{mathrm{WER}}}right)right)$$

(1)

or alternatively expressed as

$${I}_{mathrm{C}}={I}_{{rm{C}}0}+{I}_{{rm{overdrive}}}$$

(2)

Here μ_B, g, P, α and γ are the Bohr magneton, g-factor, spin polarization, Gilbert damping constant and gyromagnetic ratio, respectively. In the long-pulse limit, I_C reaches a saturation lower bound value that is given by ({I}_{{rm{C}}0}=frac{4ealpha gamma {E}_{mathrm{B}}}{{mu }_{mathrm{B}}{gP}}) and which is governed by E_B and α. The efficiency of the switching process is often characterized by the term (frac{Delta }{{I}_{{rm{C}}0}}), which should be as high as possible. (frac{Delta }{{I}_{mathrm{C0}}}) can be increased, for example, by suppression of spin pumping by using an MgO dielectric cap layer on top of the FL^40,41 or making use of two tunnel barriers⁴² to increase the spin torque. In the short-pulse limit, the switching current is increased above the saturation lower bound value by the term ({I}_{{rm{overdrive}}}=frac{4e}{{mu }_{mathrm{B}}{gP}}frac{{M}_{mathrm{s}}{V}_{{rm{m}}}}{4{t}_{mathrm{p}}}log left(frac{{pi }^{2}Delta }{4{mathrm{WER}}}right).) This term scales with ({t}_{mathrm{p}}^{-1}) and the magnetic moment of the FL.

It is clear from equation (1) that a high M_s is detrimental for current-driven switching at short pulses, although, for typical ferromagnets, it contributes towards a high E_B. By contrast, ferrimagnets that have a low M_s and high H_k, such as Mn₃Ge used here, can attain a high E_B while reducing I_overdrive. We now consider the overdrive term by studying the ratio (frac{{J}_{mathrm{C}}}{{J}_{{mathrm{C0}}}}), where J_C is the switching current density and J_C0 is the threshold saturation lower bound current density. For a given t_p, equation (1) can be rewritten as

$$frac{{J}_{mathrm{C}}}{{J}_{{rm{C}}0}}=left[1+frac{{tau }_{mathrm{D}}}{{t}_{mathrm{p}}}left(frac{1}{2}log left(frac{{pi }^{2}Delta }{4{mathrm{WER}}}right)right)right]$$

(3)

Here τ_D is the characteristic timescale of switching, which equals (frac{1}{alpha gamma {H}_{mathrm{k}}}). Calculated values of (frac{{J}_{mathrm{C}}}{{J}_{mathrm{C0}}}) are plotted in Fig. 4a as a function of H_k using equation (3). Note that as H_k is varied, M_s is correspondingly adjusted to maintain the value of ∆ to be 60. The value of α = 0.01 was chosen as it is close to the experimentally obtained α for magnetic FLs composed of Mn₃Ge⁴³ and CoFeB⁴¹. For longer pulse lengths (~10 ns), (frac{{J}_{mathrm{C}}}{{J}_{{rm{C}}0}}) is small and relatively insensitive to H_k. As clearly shown in Fig. 4a, for low H_k and short t_p, J_C is calculated to increase to more than an order of magnitude higher than J_C0, but, on the other hand, ferrimagnetic layers with a low M_s and high H_k will dramatically reduce (frac{{J}_{mathrm{C}}}{{J}_{{rm{C}}0}}) even at sub-nanosecond speeds.

**Fig. 4: Scaling of switching currents at high speeds.**

As shown in equation (3), there is a finite probability (WER) that a device will not switch with the writing pulse that is used to characterize the reliability of the writing process of an MTJ. Results for the Mn₃Ge-FL device showcased earlier in Fig. 3e are shown in Fig. 4b,c. Figure 4b shows how (frac{{J}_{mathrm{C}}}{{J}_{{rm{C}}0}}) at a higher WER of 0.5 varies for t_p ranging from 0.5 ns to 10 ns. Experimental data for both AP–P and P–AP switching processes are fitted with equation (3) to obtain J_C0 (6.05 MA cm⁻²) and τ_D (136 ps). On the basis of the fitted data, the STT switching efficiency (frac{Delta }{{I}_{{rm{c}}0}}) is ~1.37. This value is comparable to those reported in the literature for conventional ferromagnetic materials^7,44. For 1 ns long current pulses, we observe the increase in J_C over J_C0 to be ~38% (for WER = 0.5) in our device. For MTJs with CoFeB FLs, this increase is typically more than 200% (refs. ^28,45). This is due to the large difference in M_s for these two materials.

During device operations, a much lower WER is required so J_C must be increased. J_C at WER as low as 10⁻⁷ are plotted for several values of t_p in Fig. 4c. Fits to the WER versus J_C plots for each t_p are shown as solid lines (Methods). For some J_C, no errors were detected for the maximum number of switching attempts (10⁷) that we performed in our study so they are plotted at WER = 10⁻⁷ (hollow points). From these data, a plot of (frac{{J}_{mathrm{C}}}{{J}_{mathrm{C0}}}) versus t_p for WER = 10⁻⁶ is shown in Fig. 4d. The values of (frac{{J}_{mathrm{C}}}{{J}_{mathrm{C0}}}) agree well with those calculated using the value of τ_D obtained from the data in Fig. 4b. In particular, (frac{{J}_{mathrm{C}}}{{J}_{mathrm{C0}}}) for 1 ns long current pulses at WER = 10⁻⁶ is ~1.85, which is again considerably smaller than that for conventional ferromagnetic FLs²⁸. This shows the suitability of Mn₃Ge FL for achieving highly reliable switching at short pulse lengths without a notable increase in J_C.

Conclusion

In summary, we have developed CMOS-compatible MTJ non-volatile memory elements that use small switching current densities (~10 MA cm⁻²) at a short write pulse length of 0.5 ns. This is enabled by the use of a ferrimagnetic Heusler memory layer that has a very low saturation magnetization, about six times lower compared with the ferromagnetic CoFeB-based FLs used in today’s MRAM products. The low switching currents that we have demonstrated will allow for the use of minimum size drive transistors, thereby enabling a reduction in the footprint of the memory cell and overcoming a roadblock to the widespread application of spintronic technologies based on MTJ devices. The family of Heusler compounds is so extensive that one can anticipate that other members of this family will have superior performance to that presented here.

Methods

Magnetic measurements

The magnetic measurements of the Mn₃Ge film were conducted using a SQUID-VSM magnetometer (Quantum Design).

Device sizes

The film stacks were patterned into circularly shaped MTJ devices, 30–40 nm in diameter, using electron-beam lithography and Ar ion-beam milling. The physical size of the MTJs, as determined through TEM cross-sectional images, can slightly vary from the nominal size. We find that the electrical size estimated using R_P and the value of resistance-area (RA) product obtained from CIPT measurements agrees well with the size determined from TEM. Therefore, we use the electrical size to refer to the device size throughout the paper.

Film growth, stack and characterization

The samples used in these studies were prepared using an ultra-high vacuum chamber with a base pressure of ~10⁻⁹ Torr. While the Ta and Mn₃Ge layers were deposited using ion-beam deposition (with Kr gas) at a pressure 10⁻⁴ Torr, all the other layers were deposited by d.c. magnetron sputtering at an Ar sputter gas pressure of 3 mTorr (the MgO tunnel barrier was deposited by radio frequency sputtering at the same pressure). The nitride layers were deposited by reactive sputtering from a metallic target with a gas mixture of Ar and N₂. The ScN layer was deposited from a Sc target with sputter gas consisting of 85% Ar and 15% N₂. The ScN thickness can range anywhere from an interfacial layer (nominally 1 Å) up to several 100 Ås. All the films were deposited at room temperature. All MTJ stacks utilizing a CoFeB RL, in order to set PMA in the CoFeB (either by itself or in SAF structure), were in situ annealed at 300 °C after the Ta layer was deposited. Small changes in the reported annealing temperatures are related to differences in structure that may require small changes in optimal deposition and annealing conditions. As mentioned in the main text, a thick 400 Å Cr layer is included within MTJ stack to allow for CIPT measurements. The SAF structure consisted of an upper (Co/Pt) multilayer designed to have a higher moment than the lower CoFeB/Ta/(Co/Pt) multilayer of the film structure. The XRDs in Fig. 1c are normalized to the Si substrate peak. We determined the film stoichiometry by Rutherford backscattering measurements on calibration films.

WER measurements

WER measurements were performed by connecting a Keithley 2602A multimeter and a pulse generator through a bias tee to the MTJ device. The 2602A unit was used for sending the reset pulse to set the initial state of the MTJ and reading the resistance of the MTJ. Either a Picosecond Pulse Generator 10070A or a Tektronix AWG610 was used for sending the write pulse. The 10070A has a rise time of 55 ps and was used for the shorter writing pulses. Both the reset and write resistance states were measured after the state of the MTJ was set. WER was determined by first setting the MTJ to either the AP or P state using an appropriate current pulse and then applying a switching current pulse. The state of the MTJ was then read with a much smaller current sensing pulse. J_C for WER = 0.5 is obtained after accumulating the WER for different J_C and using that data to obtain the fit for J_C at WER = 0.5 (Fig. 4b). For deeper error rate measurements (from Fig. 4c), measurement at a particular J_C is run until any of these conditions are satisfied; either 10 errors are accumulated or 10⁷ trials have been performed. The WER then obtained is passed through the inverse of the standard normal cumulative distributive function and then linearly fitted against J_C (solid lines) in Fig. 4c. The smallest WER that we can measure is limited by the amount of time that it takes for a measurement to conclude. WER measurements up to 10⁻⁶ or 10⁻⁷ can be used to estimate J_C at lower WER.