Embedded non-volatile memories have great impacts on energy-efficient electronics, including Internet-of-Thing, Artificially Intelligent (AI), among others. To be successful, non-volatile memories have to satisfy several requirements, such as high writing endurance, high capacity, high speed, and low fabrication cost. Among many emerging non-volatile memory technologies, magnetoresistive random-access memory (MRAM) is one of the most promising that have gained development commitment from several leading semiconductor companies. The latest MRAM technology using the sophisticated spin–transfer–torque (STT) writing technique has just been commercially available very recently, but already found various important applications, such as highly efficient AI chips. However, in STT-MRAM, a large writing current has to be injected directly to magnetic tunneling junctions (MTJs), which leads to reliability issues such as accelerated aging of the oxide tunnel barrier1. In addition, large writing currents require large driving transistors, making it difficult to increase the bit density of STT-MRAM beyond the 1 Gbit capacity. Recently, the spin–orbit–torque (SOT) technique has emerged as a promising writing method for the next generation MRAM2. In SOT-MRAM, a charge current flowing in a non-magnetic layer with large spin–orbit interaction can generate a pure spin current by the spin Hall effect. The pure spin current is then injected to the magnetic free layer for magnetization switching. The relationships between the spin current IS and the charge current IC is given by IS = (/2e)(L/t)θSHIC, where L is the MTJ size, and t the thickness of the spin Hall layer, and θSH is the spin Hall angle. Theoretically, the charge-to-spin conversion efficiency (L/t)θSH in SOT-MRAM can be larger than unity, meaning that lower driving currents can be expected. Furthermore, since there is no current flowing into the MTJs, reliability can be significantly improved. Finally, since the spin-polarization of the pure spin current is perpendicular to the magnetization direction of the free magnetic layer, the spin torque is maximized and the magnetization can switch very fast (< ns) in SOT-MRAM with perpendicular magnetic anisotropy (PMA)3. Because of those merits of SOT-MRAM comparing with STT-MRAM, there have been huge efforts to find spin Hall materials with large θSH and high electrical conductivity for SOT-MRAM implementation. Heavy metals, such as Pt4,5, Ta2, and W6, have been studied extensively as candidates for the spin Hall layer in SOT-MRAM, since they have been used in STT-MRAM manufacturing. However, θSH of heavy metals is of the order of ~ 0.1, and the typical critical switching current density in heavy metals/ferromagnet bilayers with PMA is of the order of 107 Acm−2 to 108 Acm−2, which is too high for Si electronics7. Thus, finding a material with large θSH of the order of 10 is of emergent need.

Recently, very large θSH (> 1) has been observed in topological insulators (TIs), making them very promising for SOT-MRAM8,9. TIs are quantum materials having gaped bulk states but Dirac-like metallic surface states10,11,12 with spin-momentum locking13,14, owning to their large spin–orbit interaction and band structure topology. Although the origin of large θSH of TIs was a controversial topic due to the coexistence of the bulk current and surface current15, it has been clearly shown to originate from the surface states rather than the bulk states in carefully controlled experiments16. However, in most TI thin films, the limited surface density of states restricts the electrical conductivity σ to the order of ~ 104 Ω−1 m−1 (for example, σ ~ 5.7 × 104 Ω−1 m−1 for Bi2Se3 and 1.8 × 104 Ω−1 m−1 for (Bi0.07Sb0.93)2Te3), which results in a large shunting current when in contact with other metallic layers. Second, most TI thin films studied so far are single crystalline films grown epitaxially by the laboratory molecular beam epitaxy (MBE) technique on dedicated III-V semiconductor substrates, which is not compatible with mass production. Furthermore, ultralow power SOT magnetization switching at room-temperature with current densities of the order of 105−106 Acm−2 have been demonstrated mostly in TI/magnetic layers whose magnetic anisotropy is very small, such as in a Bi2Se3/NiFe bilayer with nearly zero in-plane magnetic anisotropy17, or in a Bi2Se3/CoTb bilayer with a small magnetization of 200 emu/cc and a PMA field of less than 2 kOe18. For TI being a practical material for SOT-MRAM, the following three minimum requirements must be satisfied: (1) a large spin Hall angle of the order of 10, (2) large electrical conductivity σ of order of 105 Ω−1 m−1, and (3) can be deposited using sputtering deposition. The first two requirements are for optimization of the switching current density and switching energy (see Suppl. Info. Section 1). The closest attempt is BixSe1-x thin films deposited by the sputtering technique, which shows a promisingly large θSH = 8.7–18.6 but with the expense of reduced electrical conductivity (σ ~ 7.8 × 103 Ω−1 m−1)19. Thus, the three requirements for a practical spin Hall material for SOT-MRAM have not yet been achieved.

Among various TI material candidates, Bi1-xSbx (0.07 < x < 0.22) is the best hope. BiSb is the first discovered three-dimensional TI with small bulk bandgap (~ 20 meV) and high bulk conductivity σ of 4−6.4 × 105 Ω−1 m−113,14,20. In thin films, the quantum size effect significantly increases the band gap of BiSb, so that the current flows mostly on the surface when the thickness reaches 10 nm21. Thanks to the multi-surface states, σ of BiSb thin films is as high as 2.5 × 105 Ω−1 m−1. Furthermore, a giant spin Hall effect with θSH ~ 52 has been observed in BiSb(012) thin films in junctions with MnGa grown on GaAs substrates by MBE22. Nevertheless, following works on sputtered polycrystalline BiSb yields a maximum θSH of only 1.2 ~ 2.4 (Refs.23,24), or even no spin Hall effect25. Therefore, it is very important to demonstrate the three requirements for sputtered BiSb for any realistic applications to SOT-based spintronic devices.

In this work, we demonstrate ultrahigh efficient SOT magnetization switching in fullly sputtered BiSb–(Co/Pt) multilayers with large PMA. We show that the sputtered BiSb has a large spin Hall angle of θSH = 10.7 and high electrical conductivity of σ = 1.5 × 105 Ω−1 m−1, thus satisfying all the three requirements for SOT-MRAM implementation. Despite the large PMA field of 5.2 kOe of the (Co/Pt) multilayers, we achieve robust SOT magnetization at a low current density of 1.5 × 106 Acm−2. Our results demonstrate the potential of BiSb topological insulator for ultralow power SOT-MRAM and other SOT-based spintronic devices.

High quality topological insulator—perpendicularly magnetized ferromagnetic multilayers

The first step in this study is to demonstrate that it is possible to use the sputtering deposition technique for fabrication of high-quality topological insulator and perpendicularly magnetized ferromagnetic multilayers. Considering the large surface roughness of TIs (~ 5–6 Angstrom for BiSb) and atomic interdiffusion during the annealing process of the MTJ, it is not realistic to deposit the MTJ with large PMA on top of the TI layer. Instead, the MTJ should be deposited and fabricated first, then the TI layer should be deposited on top of the free magnetic layer at the last step. Furthermore, to achieve high enough thermal stability, the free magnetic layer should be composed of ferromagnetic multilayers with high PMA that couple ferromagnetically or antiferromagnetically to the CoFeB layer26. Noting that the (Co/Pt)n (n = 2–6) multilayers have been frequently used in MRAM production as a part of the synthetic antiferromagnetic reference layer with large PMA for pinning, we chose the (Co 0.4 nm /Pt 0.4 nm)2 multilayers (referred below as CoPt) to realize thin ferromagnetic multilayers with large PMA for evaluating the SOT performance of the BiSb layer (see Suppl. Info. Section 2).

Figure 1a shows the studied multilayer heterostructure, which consists of perpendicularly magnetized (Co 0.4 nm/Pt 0.4 nm)2 multilayers/10 nm Bi0.8Sb0.15 topological insulator layer, capped by 1 nm MgO/1 nm Pt (not shown). The multilayers were deposited on c-plane sapphire substrates by a combination of direct current (DC) and radio-frequency (RF) magnetron sputtering in a multi-cathode chamber. Characterization by X-ray diffraction and transmission electron microscopy shows that the BiSb layer is highly textured with the dominant (110) orientation (see Suppl. Info. Section 3). We also evaluated several sputtered BiSb single layers on sapphire substrates and observed the existence of metallic surface states as well as insulating bulk states with a band gap more than 170 meV for thin (< 20 nm) BiSb films27.

Figure 1
figure 1

Perpendicularly magnetized Co/Pt ferromagnetic multilayers – topological insulator BiSb heterostructure. (a) Schematic structure of our multilayers. (b) Optical image of a Hall bar device and measurement configuration. (c) Magnetization curves of the Co/Pt multilayers. (d) Hall resistance of a Hall bar device measured with a perpendicular magnetic field.

For electrical measurements, we fabricated 25 μm-wide Hall bars by optical lithography. An optical image of a Hall bar device with the experiment setup is shown in Fig. 1b. From the parallel resistor model, we estimate that the conductivity of the BiSb layer is 1.5 × 105 Ω−1 m−1, which is much higher than that of sputtered Bi1-xSex, and close to that of MBE grown BiSb on GaAs(111)A substrates. This demonstrates that is possible to grow highly conductive BiSb topological thin films on top of perpendicularly magnetized metallic layers by the sputtering technique. Thanks to the high conductivity of the BiSb layer, 50% of the applied current flows into the BiSb and contribute to the SOT magnetization switching.

Figure 1c shows the magnetic hysteresis curves of the as-grown sample measured by a superconducting quantum interference device (SQUID) under an in-plane and out-of-plane external magnetic field, respectively. The saturation magnetization MCoPt, normalised by the CoPt layer thickness (tCoPt = 1.6 nm), is 613 emucm−3. The uniaxial anisotropy field Hk = 15 kOe is very large for the as-grown film. After Hall bar device fabrication by optical lithography undergoing several circles of thermal annealing, Hk is reduced to 5.2 kOe (see Suppl. Info. Section 4), but is still much larger than that of NiFe or CoTb used in previous works. This CoPt layer yields a thermal stability factor for magnetization switching of Δ = 38, similar to that of the CoFeB(/MgO) free layer of perpendicular MTJs with diameters of 20–40 nm28,29. Figure 1d shows the anomalous Hall resistance (RH) measured for a Hall bar under a sweeping out-of-plane field, which confirms PMA of the CoPt layer.

Evaluation of the spin Hall angle by the second harmonic Hall measurements

Next, we performed the second harmonic Hall measurements to evaluate the spin Hall angle30. An alternating current (AC) J = J0sin ωt (ω = 259.68 Hz) was applied to the Hall bar under a sweeping external field along the x direction. We measured the 2nd harmonic Hall resistance \(R_{\text{H}}^{{2{\upomega }}}\), which is originated from the oscillation of the net magnetic moment under the spin–orbit effective magnetic fields31,32,33. Figure 2a shows a representative \(R_{\text{H}}^{{2{\upomega }}}\)- Hx curve measured at JBiSb = 3.6 × 105 Acm−2. The second harmonic Hall resistance \(R_{\text{H}}^{{2{\upomega }}}\) can be expressed as33

$$R_{\text{H}}^{{2{\upomega }}} = \frac{{R_{{\text{H}}} }}{{2}}\frac{{H_{{{\text{AD}}}} }}{{H_{x} - H_{{\text{k}}} (H_{x} /\left| {H_{x} } \right|)}} + R_{{{\text{PHE}}}} \frac{{H_{{{\text{FL}} + {\text{Oe}}}} }}{{H_{{x}} }} + R_{{{\text{thermal}}}} \frac{{H_{x} }}{{{|}H_{x} {|}}}$$

where HAD is the antidamping-like effective field, HFL+Oe is the sum of the fieldlike and Oesterd field, RPHE is the planar Hall resistance, and Rthermal is the contribution from the anomalous Nernst (ANE) and spin Seebeck (SSE) effects. Note that the contribution of the fieldlike and Oesterd field to \(R_{\text{H}}^{{2{\upomega }}}\) is much smaller than that of the antidampinglike field in BiSb34,35. Fitting Eq. (1) to the high field data in the \(R_{\text{H}}^{{2{\upomega }}}\)- Hx curve yields HAD (red curves in Fig. 2a). Figure 2b shows HAD as a function of JBiSb. From the HAD / JBiSb gradient, we can calculate the effective spin Hall angle \({\theta }_{\mathrm{SH}}^{\mathrm{eff}}\) = \(\frac{2e}{\hbar }M_{{{\text{CoPt}}}} t_{{{\text{CoPt}}}} \frac{{H_{{{\text{AD}}}} }}{{J^{{{\text{BiSb}}}} }}\) = 12.3, where e is the electron charge and \(\hbar\) is the reduced Plank constant. Recently, it was reported that CoPt multilayers can generate a “self” spin–orbit torque36. We found that the “self” spin–orbit torque in CoPt can contribute to 13% of the total spin–orbit torque (see Suppl. Info. Section 6). Furthermore, we confirmed that there is no artifact contribution from the asymmetric magnon scattering mechanism (see Suppl. Info. Section 7), as observed in the case of the magnetic topological insulator (CrBiSb)2Te337. Subtracting the contribution from CoPt, we obtain the intrinsic spin Hall angle of BiSb θSH = 10.7, which demonstrates the feasibility of BiSb for ultralow power SOT-MRAM.

Figure 2
figure 2

Evaluation of the spin Hall angle by the second harmonic measurements. (a) 2nd harmonic Hall resistance as a function of the in-plane external magnetic field H applied along the x direction. The red curves are the theoretical fitting using Eq. (1). (b) HAD as a function of JBiSb.

Ultrahigh efficient spin–orbit torque magnetization switching by DC and pulse currents

Next, we demonstrate ultrahigh efficient and robust SOT magnetization switching in the CoPt/BiSb multilayers. Figure 3 shows the SOT magnetization switching by DC currents with an applied external field along the x direction. We achieved Hall resistance switching whose amplitude is consistent with that of the Hall resistance loop shown in Fig. 1d, indicating full magnetization switching. The switching direction is reversed when the external magnetic field direction is reversed, which is consistent with the characteristic of SOT. Typical DC threshold switching current density \(J_{{{\text{th}}}}^{{{\text{BiSb}}}}\) is 1.5 × 106 Acm−2 at the bias field of 2.75 kOe. Note that thanks to the high electrical conductivity σ = 1.5 × 105 Ω−1 m−1 of BiSb, the total current density including the shutting current in the CoPt is kept at 2.6 × 106 Acm−2. Next, we performed SOT magnetization switching by pulse currents. Figure 4a,b show representative SOT switching loops by 0.1 ms pulse currents at + 1.83 kOe and − 1.83 kOe, respectively. Figure 4c plots \(J_{{{\text{th}}}}^{{{\text{BiSb}}}}\) at various pulse width tpulse, and the theoretical fitting using the thermal activation model \(J_{{{\text{th}}}}^{{{\text{BiSb}}}}\) = \(J_{{0}}^{{{\text{BiSb}}}}\)×\(\left[ {1 - \frac{1}{\Delta }\ln \left( {\frac{{\tau_{{{\text{pulse}}}} }}{{\tau_{0} }}} \right)} \right]\)38, where \(J_{{0}}^{{{\text{BiSb}}}}\) is the zero-kelvin threshold switching current density, ∆ is the thermal stability factor, and 1/τ0 = 1 GHz (τ0 = 1 ns) is the attempt switching frequency. The fitting yields \(J_{{0}}^{{{\text{BiSb}}}}\) = 4.6 × 106 Acm−2 and Δ = 38. Because magnetization switching occurs by domain wall nucleation and domain wall motion, Δ reflects the energy barrier of the volume with size equal to the domain wall width, i.e. Δ should be considered as the energy barrier to nucleate a domain wall, rather than the energy barrier for coherently switching of the whole volume of the magnetic layer39. Therefore, Δ evaluated by this way is smaller than that should be expected for switching the whole volume of the magnetic layer. Nevertheless, the obtained Δ of CoPt is large enough to ensure that the total Δ in ferromagnetically (antiferromagnetically) coupled CoFeB/Ta(Ru)/CoPt free layer can exceeds 60 for 10 years thermal stability, while the switching current density remains the same40. Finally, we demonstrate robust SOT switching in the CoPt/BiSb junction. For this purpose, we applied a sequence of 150 pulses (\(J_{{{\text{th}}}}^{{{\text{BiSb}}}}\) =  ± 4.4 × 106 Acm−2, tpulse = 0.1 ms) as shown in the top panel of Fig. 4c. The Hall resistance data recorded for a total of 150 pulses under ± 1.83 kOe are shown in the bottom panel in Fig. 4c. We observed a robust SOT switching with no change in the device characteristics, indicating that the BiSb topological insulator deposited by the sputtering technique has great potential for realistic SOT-MRAM.

Figure 3
figure 3

SOT magnetization switching by DC currents. Switching loops measured under an in-plane magnetic field applied along (a) + x direction and (b) − x direction.

Figure 4
figure 4

SOT magnetization switching by pulse currents. (a,b) Switching loop by 0.1 ms pulse currents under an in-plane magnetic field of H =  + 1.83 kOe and − 1.83 kOe, respectively. (c) Threshold current density \(J_{{{\text{th}}}}^{{{\text{BiSb}}}}\) as a function of tpulse. (d) Robust SOT magnetization switching by 0.1 ms pulse current.

SOT performance benchmarking and future prospective

Table 1 summarizes θSH, σ, the spin Hall conductivity σSH = (ħ/2e)σθSH, and the SOT normalized power consumption Pn at room temperature of several heavy metals and TIs. Here, θSH of TIs are their best values reported in literature. For the calculation of the Pn, we assumed bilayers of spin Hall material (thickness t = 6 nm for heavy metals and t = 10 nm for TIs) and CoFeB (thickness tFM = 1.5 nm, conductivity σFM = 6 × 105 Ω−1 m−1). Considering the shunting current in the ferromagnetic layer, the SOT power consumption is proportional to (σt + σFMtFM)/(σtθSH)2. One can see that not only θSH but also σ affect the SOT power consumption, a fact usually overlooked in literature. For example, while the sputtered BixSe1-x has a much larger spin Hall angle (θSH = 18.6) than that (θSH = 3.5) of MBE-grown Bi2Se3, their power consumption is nearly the same, because BixSe1-x has poorer crystal quality than Bi2Se3 and thus very low conductivity. Meanwhile, the sputtered BiSb thin film in this work shows both high σ = 1.5 × 105 Ω−1 m−1 and large θSH = 10.7, which are optimal for both small switching current density and small switching power consumption41. Indeed, the switching power consumption for sputtered BiSb is 50 times smaller than that for sputtered BixSe1-x, and over 300 times smaller than that for W, which is the most used heavy metal in SOT-MRAM development. The small switching current density and switching power also help suppress failure of the spin Hall layer due to electromigration and Joule heating42. Our results demonstrate the feasibility of BiSb for not only ultralow power SOT-MRAM but also other SOT-based spintronic devices, such as race-track memories43 and spin Hall oscillators44,45.

Table 1 Spin Hall angle θSH, electrical conductivity σ, spin Hall conductivity σSH, and SOT normalized power consumption Pn of several heavy metals and topological insulators.

Finally, we discuss about the remaining challenges for realization of ultralow power BiSb-based SOT-MRAM. One of material challenges is the large surface roughness of BiSb due to the unusually large crystal grain size comparing with the layer thickness. Furthermore, atomic interdiffusion between BiSb and the free magnetic layer during annealing process of the MTJ is also a challenge to be resolved. Thereby, a realistic pathway to integrate BiSb to SOT-MRAM is to fabricate full stack MTJs with PMA (p-MTJ) first, then deposit the BiSb layer on top of the MTJs at the last step after the annealing process, i.e. BiSb-on-MTJ structure. This will help avoid the interdiffusion and the BiSb surface roughness problems in the MTJ-on-BiSb structure.


Material growth

We deposited multilayers of (0.4 nm Co / 0.4 nm Pt)2 / 10 nm Bi0.85Sb0.15 / 1 nm MgO / 1 nm Pt on c-plane sapphire substrates by DC (for Co, Pt, BiSb) and RF (for MgO) magnetron sputtering in a multi-cathode chamber. All layers are deposited by sputtering from their single targets using Ar plasma without breaking the vacuum at room temperature.

Device fabrication

The samples were patterned into 90 μm-long × 25 μm-wide Hall bars by optical lithography and lift-off. A 45 nm-thick Pt were deposited as electrodes by DC magnetron sputtering, which reduces the effective length of the devices to 50 μm.

SOT characterization

The samples were mounted inside a vacuumed cryostat equipped with an electromagnet. For the second harmonic measurements, a NF LI5650 lock-in amplifier was employed to detect the first and the second harmonic Hall voltages under sine wave excitation generated by a Keithley 6221 AC/DC current source. For the DC (pulse) current-induced SOT magnetization switching, a Keithley 2400 SourceMeter (6221 AC/DC current source) was used, and the Hall signal was measured by a Keithley 2182A NanoVoltmeter.