Nanometre-thick Si/Al Gradient Materials for Sustainable Spin Current Generation

Green materials for electron spin manipulation are essential for common spintronic applications. Recent studies have documented the efficient generation of spin torque using spin-orbit interactions (SOIs); however, SOI use relies on the employment of rare metals such as platinum. Here, we demonstrate that a nanometre-thick gradient from silicon to aluminium, which consists of readily available elements from earth resources, can produce a spin torque over three times as large as that of platinum, despite the weak SOI of these compositions. The spin-torque efficiency can be improved by decreasing the thickness of the gradient, while a sharp interface was not found to increase the spin torque. Moreover, the electric conductivity of the gradient material can be up to twice as large as that of platinum, which provides a way to reduce Joule heating losses and signal delays in spintronic device circuits. Our findings represent a pathway for sustainable spin manipulation. Non-equilibrium spin polarisation in conduction electrons enables the magnetisation direction to be manipulated via spin torque using spin-orbit interactions (SOIs) in spintronic devices such as magnetic random-access memory (MRAM) 2 and spin torque nanooscillators. Phenomena such as the spin-Hall effect (SHE) and the Rashba-Edelstein effect (REE) are commonly used to generate spin-polarised flow, i.e. spin currents (SC), in strong SOI materials. Such non-equilibrium spin polarisation generally requires specific materials with a strong SOI, usually consisting of heavy 5d metal elements such as tantalum (Ta), tungsten (W) and platinum (Pt). The figure of merit for SC generation capability is generally determined as the product of the spin Hall angle and the electric conductivity of the material. Namely, higher conductivity materials exhibit superior SC generation, while stronger SOI elements generally exhibit smaller conductivities. This has been observed in hitherto promising candidate materials for generating spin torque in MRAMs, such as β−Ta and β−W. The material choice dilemma for SC generation creates a bottleneck in spintronic device production. Poor conductivity in SC circuits can lead to other serious problems such as wiring delays and Joule losses in integrated circuits. Moreover, semiconductor device performance is often degraded by contamination with strong SOI elements. Therefore, a SOI-free technology is important for enabling SC circuit integration in electric devices. The problem of rare-metal element scarcity also poses a serious challenge to sustainable development. A promising technology for producing flow of spin angular momentum without using SOI is spin separation produced by a magnetic field gradient. Stern and Gerlach achieved such separation in 1922 with a spin current generated along a field gradient (z-axis in Fig. 1a) . This mechanism can be used to generate a spin current without using SOI. Indeed, it has been experimentally demonstrated that an emergent magnetic field can be produced based on the gyromagnetic effect on spins in metallic liquid, thin copper films, and quark-gluon plasma . A similar mechanism was recently proposed in a surface-oxidised copper film, where an emergent field was generated by a non-uniform electric current, although its physical origin is still controversial. Because a conductivity gradient leads to a vortical electric current that acts on spins as an emergent magnetic field (Fig. 1b), the magnitude of the field should depend on the thickness of electron conductivity gradient in oxides. However, the thickness of the oxidation gradient is difficult to control atomically. This poses a problem not only for practical applications of spintronic devices but also for confirming the existence of emergent magnetic fields due to electric current vorticity. In this study, we successfully fabricated a nanometre-thick artificial gradient from silicon (Si) to aluminium (Al), which constitutes the secondand third-most abundant components in the Earth’s crust (Fig. 1c), and we demonstrate its ability to generate a spin torque of greater magnitude than Pt. The magnitude of the spin torque increases when the compositional gradient steepens; however, an atomically sharp interface does not significantly increase the spin torque. Moreover, as reported in the study of a surface-oxidised copper (Cu) film, we observed a large non-reciprocity that is regarded as evidence for a SC generated via an emergent magnetic field due to electric current vorticity. Fabrication and structural analysis of Si/Al gradient materials We fabricated Si(10)/Al(ti/2)/ Si(ti/2)/Al(10)/Ni95Cu5(10)/SiO2(20) (unit: nm) multilayer strips on a thermally oxidised Si substrate by means of a conventional lift-off method using magnetron sputtering and photolithography. Here, ti is the thickness of the interfacial Al/Si insertion, which varied from 0.5 to 2.0 nm at intervals of 0.5 nm. Al is a very conductive metal with an open 2p shell. Although Si is adjacent to Al in the periodic table, it is a semiconductor with much lower electrical conductivity than Al. Sputter deposition processes generally lead to atomic or metallographic disturbances at an interface due to the large kinetic energy of the sputtered particles. An insertion of a few nm of Al/Si therefore increases the mixed region at the interface between 10-nm-thick Si and Al layers. Indeed, microstructural analysis using high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) revealed an increase in the thickness of the compositional gradient from Si to Al with increasing ti. Figure 2a,b displays cross-sectional HAADF-STEM images near the Si/Al interface with ti = 2.0 and 1.0 nm, respectively. In comparison, a cross-section of a Si/Al interface without Al/Si insertion (i.e. ti = 0 nm) is also shown in Fig. 2c, in which a sharp interface was clearly observed. The Al/Si insertion (with thickness ti) obscures the Si/Al boundary. Moreover, aggregation of Al and/or Si was observed at the Si/Al interface for ti = 2.0 nm. The formation of such aggregation features is attributed to non-solid solution atomic mixing between Si and Al. Energy dispersive spectroscopy (EDS) line profiles of each element (Si and Al) were obtained by averaging the signals in the dashed box area shown in Fig. 2d,e,f which explicitly confirm the formation of a Si/Al gradient with a transition thickness that systematically varies with ti. The transparent bold lines in Fig. 2g,h,i show the best fit to the following equation: Composition = C1 +C2 2 ± C2 −C1 2 tanh (

Non-equilibrium spin polarisation in conduction electrons enables the magnetisation direction to be manipulated via spin torque using spin-orbit interactions (SOIs) in spintronic devices such as magnetic random-access memory (MRAM) 1, 2 and spin torque nanooscillators [3][4][5] . Phenomena such as the spin-Hall effect (SHE) [6][7][8][9][10] and the Rashba-Edelstein effect (REE) [11][12][13][14] are commonly used to generate spin-polarised flow, i.e. spin currents (SC), in strong SOI materials. Such non-equilibrium spin polarisation generally requires specific materials with a strong SOI, usually consisting of heavy 5d metal elements such as tantalum (Ta), tungsten (W) and platinum (Pt). The figure of merit for SC generation capability is generally determined as the product of the spin Hall angle and the electric conductivity of the material. Namely, higher conductivity materials exhibit superior SC generation, while stronger SOI elements generally exhibit smaller conductivities. This has been observed in hitherto promising candidate materials for generating spin torque in MRAMs, such as β −Ta and β −W [15][16][17][18][19] . The material choice dilemma for SC generation creates a bottleneck in spintronic device production. Poor conductivity in SC circuits can lead to other serious problems such as wiring delays and Joule losses in integrated circuits. Moreover, semiconductor device performance is often degraded by contamination with strong SOI elements. Therefore, a SOI-free technology is important for enabling SC circuit integration in electric devices. The problem of rare-metal element scarcity also poses a serious challenge to sustainable development.
A promising technology for producing flow of spin angular momentum without using SOI is spin separation produced by a magnetic field gradient. Stern and Gerlach achieved such separation in 1922 with a spin current generated along a field gradient (z-axis in Fig. 1a) 20,21 . This mechanism can be used to generate a spin current without using SOI. Indeed, it has been experimentally demonstrated that an emergent magnetic field can be produced based on the gyromagnetic effect on spins in metallic liquid 22 , thin copper films [23][24][25] , and quark-gluon plasma 26,27 . A similar mechanism was recently proposed in a surface-oxidised copper film 28 , where an emergent field was generated by a non-uniform electric current, although its physical origin is still controversial [29][30][31][32] . Because a conductivity gradient leads to a vortical electric current that acts on spins as an emergent magnetic field (Fig. 1b), the magnitude of the field should depend on the thickness of electron conductivity gradient in oxides. However, the thickness of the oxidation gradient is difficult to control atomically. This poses a problem not only for practical applications of spintronic devices but also for confirming the existence of emergent magnetic fields due to electric current vorticity.
In this study, we successfully fabricated a nanometre-thick artificial gradient from silicon (Si) to aluminium (Al), which constitutes the second-and third-most abundant components in the Earth's crust (Fig. 1c), and we demonstrate its ability to generate a spin torque of greater magnitude than Pt. The magnitude of the spin torque increases when the compositional gradient steepens; however, an atomically sharp interface does not significantly increase the spin torque. Moreover, as reported in the study of a surface-oxidised copper (Cu) film 28 , we observed a large non-reciprocity that is regarded as evidence for a SC generated via an emergent magnetic field due to electric current vorticity.

Fabrication and structural analysis of Si/Al gradient materials
We fabricated Si(10)/Al(t i /2)/ Si(t i /2)/Al(10)/Ni 95 Cu 5 (10)/SiO 2 (20) (unit: nm) multilayer strips on a thermally oxidised Si substrate by means of a conventional lift-off method using magnetron sputtering and photolithography. Here, t i is the thickness of the interfacial Al/Si insertion, which varied from 0.5 to 2.0 nm at intervals of 0.5 nm. Al is a very conductive metal with an open 2p shell. Although Si is adjacent to Al in the periodic table, it is a semiconductor with much lower electrical conductivity than Al. Sputter deposition processes generally lead to atomic or metallographic disturbances at an interface due to the large kinetic energy of the sputtered particles. An insertion of a few nm of Al/Si therefore increases the mixed region at the interface between 10-nm-thick Si and Al layers. Indeed, microstructural analysis using high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) revealed an increase in the thickness of the compositional gradient from Si to Al with increasing t i . Figure 2a,b displays cross-sectional HAADF-STEM images near the Si/Al interface with t i = 2.0 and 1.0 nm, respectively. In comparison, a cross-section of a Si/Al interface without Al/Si insertion (i.e. t i = 0 nm) is also shown in Fig. 2c, in which a sharp interface was clearly observed. The Al/Si insertion (with thickness t i ) obscures the Si/Al boundary. Moreover, aggregation of Al and/or Si was observed at the Si/Al interface for t i = 2.0 nm. The formation of such aggregation features is attributed to non-solid solution atomic mixing between Si and Al. Energy dispersive spectroscopy (EDS) line profiles of each element (Si and Al) were obtained by averaging the signals in the dashed box area shown in Fig.  2d,e,f which explicitly confirm the formation of a Si/Al gradient with a transition thickness that systematically varies with t i . The transparent bold lines in Fig. 2g,h,i show the best fit to the following equation: where C 1 and C 2 are the composition at each end of an EDS scan, z int is the centre position of the interface, and L is the thickness of the compositional gradient from Si to Al. The fit of Eq.

Electric current-induced spin torque
When an electric current is applied to the bilayer strip consisting of a nonmagnet (NM) and a ferromagnet (FM), part of the SC generated in the NM is transmitted towards the FM followed by a spin torque applied on the magnetisation of the FM. This torque is generally referred to as a damping-like (DL) torque, τ τ τ DL , whereas a separate spin torque known as the field-like (FL) torque, τ τ τ FL , arises from the SC reflected at the NM/FM interface. The effect of these torques on a magnetisation m m m produced by a SC with a polarisation σ σ σ s is described by the following equation: where ξ DL and ξ FL are the efficiencies of the DL and FL torques with respect to an electric current density j c , while e, µ 0 , and ℏ are the elementary charge, permeability of a vacuum, and the reduced Planck's constant, respectively. M s and d FM are the saturation magnetisation and FM thickness, respectively. Both these orthogonal torques play an important role in magnetisation switching, which has been widely investigated in applications for non-volatile magnetoresistive memory and magnetic logic devices.
To evaluate the strength of the spin torque produced by applying an electric current to the sample, we conducted spin-torque ferromagnetic resonance (ST-FMR) measurements 15,16,18,[33][34][35] . The theory of this technique is described in Supplementary note S1. In the ST-FMR experiment, an alternating current was applied to the strip using a microwave with amplitude of 20 dBm and a frequency of 20 GHz. The direct current voltage V dc magnitude due to the ST-FMR excitation was measured while sweeping the external field from 0 to 2.0 T. Figure 3a shows a typical ST-FMR spectrum for t i = 0.5 nm, which can be reproduced by a combination of symmetric-and antisymmetric-Lorentzian functions, as shown in Fig. 3b. From the amplitude ratio between the symmetric-and antisymmetric-Lorenzian components, V s /V a , we can evaluate the spin torque efficiency ξ FMR , which is commonly used as a rough estimate of ξ DL when ξ FL is negligible in ST-FMR. From Eq. (S1) in Supplementary note S1 and the value of V s /V a for the ST-FMR spectrum at a given resonant field B r , we obtain ξ FMR = 0.029 ± 0.004 for a sample with t i = 0.5 nm, whose magnitude is about half the value for a Pt(10 nm)/Ni 95 Cu 5 (10 nm) bilayer film (ξ FMR = 0.066 ± 0.005). Figure 3c shows ξ FMR as a function of t i . As shown in Fig. 3c, a lower t i leads to a larger ξ FMR , although the ξ FMR for t i = 0 nm is similar to the value for a reference sample with an Al(10 nm)/Ni 95 Cu 5 (10 nm) structure fabricated on a thermally oxidised Si substrate. This result implies that most of the SC is produced in the 10-nm-thick Al region via SHE and that a sharp Si/Al interface does not result in additional SC generation. That is, the nanometre-thick gradient from Si to Al is the most important factor for SC generation in the sample.
In general, a large difference in electric conductance between NM and FM means that the contribution of ξ FL cannot be ignored. As shown in Supplementary note S3, both ξ DL and ξ FL can be determined from the dependence of ξ FMR on d FM . Furthermore, the ratio between ξ DL and ξ FL generally depends on the condition of a NM/FM interface, i.e. an Al/Ni 95 Cu 5 interface. In other words, the value of ξ FL /ξ DL is independently determined with respect to the thickness of the compositional gradient at the Si/Al interface. Figure 3d contains a plot of ξ FMR as a function of ξ DL and ξ FL that is expected from Eq.(S2) of Supplementary note S3. From the ξ DL and ξ FL values for t i = 1.0 nm in Supplementary note S3, ξ FL /ξ DL was determined to be 3 for the Al/Ni 95 Cu 5 interface. The dashed line in Fig. 3d indicates the condition ξ FL /ξ DL = 3. The solid circles on the dashed line in Fig. 3d indicate the experimental ξ FMR values obtained in Fig. 3c. The x-and y-axes values for each circle in Fig.  3d therefore represent ξ DL and ξ FL of each t i , respectively. As shown in Fig. 3d, the magnitudes of both ξ DL and ξ FL clearly increase with decreasing t i , which is the thickness of the compositional gradient from Si to Al. Surprisingly, the ξ DL value for t i = 1.0 nm reaches a magnitude three times larger than that of Pt; further increase in ξ DL would be expected for t i = 0.5 nm, even though only weak SOI elements, i.e. Si and Al, were used for the NM. The ratio between ξ DL and ξ FL suggests that the reflected SC is three times larger than the transmitted SC. This is attributed to the fact that electric conductivity in the SC absorber, the Ni 95 Cu 5 layer, is an order of magnitude smaller than that of the SC injector, or the Al layer. Further increases in ξ DL can therefore be realised using ferromagnetic materials with a higher electrical conductivity than Ni 95 Cu 5 .

Origin of spin torque accompanied by Si/Al gradient materials
Below we discuss the origin of spin torque in the Si/Al/Ni 95 Cu 5 trilayer film. Similarly to a torque generated via SHE and/or REE with an electric current along the x-axis, the variations of V s and V a with respect to the direction of the external magnetic field can be explained by assuming that the electron spin is polarised along the y-axis (see Supplementary note S4) [36][37][38] . In the Si/Al/Ni 95 Cu 5 trilayer film, most of the electric current flows not within the Si or Ni 95 Cu 5 layer, but in the Al layer because the electric conductivity of Al is much larger. However, the spin torque produced by the electric current flow in the Al layer via SHE can be ignored because the spin Hall angle for bulk Al is only 0.02 8,39 , which is 10 times smaller than the value of ξ DL in our sample with t i = 1.0 nm. Inversion symmetry breaking at the Al/Ni 95 Cu 5 interface can also produce a spin torque. However, this effect should be independent of t i because the atomic and/or structural asymmetry condition at the Al/Ni 95 Cu 5 interface is determined by the sequential sputtering processes of Al and Ni 95 Cu 5 , which are commonly used in all samples. Consequently, the dependence of ξ DL on t i is not attributed to a spin torque caused by either bulk or interface SOI. An inversion symmetry breaking leading to an orbital angular momentum (OAM) flow is also enhanced at the Si/Al interface by decreasing the transition width and must be strongest at t i = 0 nm, while the smallest ξ DL was observed at t i = 0 nm, as shown in Fig. 3d. Therefore, generation of OAM flow is not the primary cause of the increase in ξ DL for the compositional gradient from Si to Al. Similarly, interface-generated SCs caused by either spin-orbit filtering or spin-orbit precession 40 can also be excluded.
An additional possible explanation for the increase in ξ DL is SC generation from electric current vorticity via spin vorticity coupling (SVC) 22,23,41,42 . A non-uniform distribution of electric current arises in the nanometre-thick Si/Al gradient from the large difference in electrical conductivity between Si and Al. Similarly to the inversion symmetry breaking, the degree of electric current non-uniformity increases with decreasing compositional gradient thickness. Conversely, specular electron scattering, which occurs at sharp interfaces, would create a uniform electric current distribution near the interface. Consequently, the electric current vorticity will be maximised when compositional gradient thickness is comparable to the effective mean free path of the electron. The optimisation of ξ DL at a certain compositional gradient thickness seems to support the SVC mechanism. Indeed, as shown in Supplementary note S5, the ratio of ξ DL increases with respect to the compositional gradient thickness by a similar order of magnitude to the value obtained by evaluating a simple model based on SVC theory. Fluctuations in the compositional gradient that disrupt the electric current vorticity are likely to be small because the strength of ξ DL clearly scales with the averaged compositional gradient thickness, as shown in Fig. 3d.

Highly non-reciprocal SC generation in Si/Al gradient materials
A highly non-reciprocal conversion between electric current and SC can be considered additional evidence that SVC contributes to SC generation. In the case of typical SOI-based phenomena, both intrinsic and extrinsic spin-dependent scattering sources exist in the material regardless of the applied electric current. On the contrary, the vorticity gradient needed for SC generation via SVC appears only when a net drift velocity exists. Namely, the SC is not accompanied by a net drift velocity and, consequently, cannot be converted into electric current.
To evaluate the conversion efficiency from a SC to electric current, as shown in Supplementary note S6, we measured the inverse SHE due to the SC 43,44 , which was produced by applying an alternating magnetic field [45][46][47][48] . Figure 4a shows the conversion efficiency θ j s → j c of the SC to an electric current as a function of ξ DL , which is proportional to the conversion efficiency θ j c → j s from the charge current to the SC. In general, an SC source with a two-dimensional geometry, such as an interfacial SOI, suppresses θ j s → j c rather than θ j c → j s although the value of θ j c → j s /θ j s → j c cannot exceed 10 49,50 . If the SC source does produce such a geometrical effect in the Si/Al gradient material, the non-reciprocity is expected to be intermediate between that of a bulk and an interfacial SOI because of its fuzzy interface. However, as shown in Fig. 4a, the sample with t i = 0.5 nm exhibits a ξ DL /θ j s → j c value much larger than 10. On the other hand, ξ DL /θ j s → j c for t i = 0 nm is unity, at which the highest two-dimensional geometry is expected. The fact that strong non-reciprocity is observed, not for t i = 0 nm but for t i = 0.5 nm, therefore suggests that the SC is generated by neither interfacial SOI nor inversion symmetry breaking at the Si/Al interface, but by angular momentum conversion from macroscopic vorticity in the electric current at the compositional gradient.

Figure of merit for spin torque switching capability of Si/Al gradient materials
To suppress the power-supply voltage for spin torque switching, the product of ξ DL and the electrical conductivity σ e of the NM must increase. Moreover, the magnitude of σ e itself should increase to reduce resistive-capacitive delays, which hinder the high-speed operation of integrated circuits. The relative abundance of elements P n is also an important parameter for sustainable development of SC-generating materials. To satisfy these demands, a figure of merit (FOM) for SC-generating materials can be determined as the product of these values as follows: From the statistical data for P n in Fig. 1a, P Si/Al = P Si × P Al = 2.24 × 10 −2 , which is seven orders of magnitude larger than the value for Pt (P Pt of 5×10 −9 ). Si/Al gradient materials therefore offer a highly superior solution for sustainable materials development. Figure 4b displays a double logarithmic graph with respect to σ e and ξ DL σ e for the Si/Al gradient samples. The data for a Pt/Ni 95 Cu 5 bilayer film are also plotted as a closed square for comparison. σ e is shown as a function of t i in Supplementary note S8. The dashed lines in Fig. 4b represent contours of ξ DL σ 2 e . In addition to ξ DL σ e , we can also increase σ e in our Si/Al gradient materials by decreaseing t i from 2.0 to 0.5 nm. This is surprising because strong SOI materials such as Pt, W, and Ta, typically exhibit low σ e . More specifically, ξ DL σ 2 e can be larger than the typical value for strong-SOI materials even at the same value of ξ DL . Indeed, we achieved a ξ FMR σ 2 e value of 14.3 (MS/m) 2 for t i = 0.5 nm, which is seven times larger than that in Pt. Moreover, the FOM for this material was 3.2×10 11 (S/m) 2 , which is seven orders of magnitude greater than that of Pt (1.0×10 4 (S/m) 2 ). The overwhelmingly higher FOM for our fabricated films demonstrates the high potential for Si/Al gradient materials to be widely used in low-energy consumption and sustainable spin devices. Fig. 1 | Green and rare-metal-free composition gradient material for spin current generation. a, Schematic illustration of the Stern-Gerlach experiment. Quantisation of spin angular momentum σ σ σ s was observed when a beam of silver atoms was split in two as it passed through a gradient of a static magnetic field B B B. b, Schematic image of a fabricated Si/Al gradient material. The Cartesian coordinate system used throughout this article is also shown in (b). A nanometre-thick compositional gradient from Si to Al exists along the z-axis with much conductivity in the Al layer than the Si layer. When an electric current j j j c is applied along the x-axis, an electric current vorticity ω ω ω = ∇ × j j j c appears, whose vector points along the y-axis. The electron spin σ σ σ s in the gradient material is polarised along the y-axis by an emergent magnetic field due to the vortical electric current. As a consequence, the gradient of ω ω ω leads to a density gradient in the non-equivalent electron spin, which produces a spin current along the z-axis. c, A plot of relative elemental abundance in the Earth's crust. Si and Al represent the most abundant semiconductor and metal materials, respectively, while typical large spin current source materials such as Ta, W, and Pt are very rare.  The circles show data for samples with various t i , whereas the square represents values for a Pt(10 nm)/Ni 95 Cu 5 (10 nm) bilayer film. The vertical bars indicate the standard deviation calculated from the leaset-squares deviation of the fitting parameters used to calculate θ j s → j c . b, Double logarithmic plot of electric conductivity σ e and ξ DL σ e for samples with t i ranging from 0 to 2.0 nm. The data for a Pt(10 nm)/Ni 95 Cu 5 (10 nm) bilayer film is also plotted for comparison(square). Dashed lines are contours of ξ DL σ 2 e .

Sample Preparation
The films were fabricated on thermally oxidised silicon substrates by magnetron sputtering at room temperature. The chamber base pressure prior to deposition was less than 5.0 × 10 −4 Pa. The deposition pressure was 0.22 Pa with an argon (Ar) flow rate of 4.0 sccm. The Al layer creation used radio-frequency (RF) deposition at 13.56 MHz with a power density of 1.4 W/m 2 and a deposition rate of 0.043 nm/s from a 99.9%-pure Al target. The Si layer was deposited by RF-sputtering with a power density of 3.5 W/m 2 and a deposition rate of 0.062 nm/s from a non-doped Si target. The Ni 95 Cu 5 layer was deposited by direct current (DC) sputtering with a power density of 1.4 W/m 2 and a deposition rate of 0.2 nm/s from a 99.9%-pure Ni 95 Cu 5 alloy target. The SiO 2 capping layer was deposited by RF-sputtering with a power density of 3.5 W/m 2 and a deposition rate of 0.044 nm/s from a 99.99%-pure SiO 2 target. The thin films were patterned into strips 10-µm-wide and 100-µm-long by photolithography and lift-off processes. An electrically shorted coplanar waveguide made from 70-nm-thick Au was connected to both ends of the strip to conduct the ST-FMR measurements. All measurements were conducted at room temperature.

Electrical Measurements
For ST-FMR measurements, we applied 20-dBm continuous sinusoidal signals with a frequency of 20 GHz in the longitudinal direction (x-axis) of the film by a signal generator. An in-plane external magnetic field was applied with an amplitude ranging from 0 to 2.0 T at a fixed angle of π/4 with respect to the x-axis. We then measured the rectified DC voltage V dc by a nanovoltmeter from a DC port of bias tee. The resulting spectra were fit by the expression V dc = V s f s (B) + V a f a (B), where f s (B) and f a (B) are the symmetric-and antisymmetric-Lorentzian functions, respectively. The inverse SHE measurement used a general spin-pumping experiment setup. We applied a continuous sinusoidal signal with amplitude of 20 dBm and a frequency of 5 GHz into the coplanar waveguide fabricated on the sample, which was patterned into a Hall-bar shape. The sample and coplanar waveguide were insulated by the insertion of 120-nm-thick SiO 2 film. The external magnetic fields with amplitudes ranging from 0 to 2.0 T were applied in the x-y plane at an angle ranging between 0 and 2π from the x-axis. The DC inverse spin Hall voltage was measured using a nanovoltmeter.

Electron microscopy characterisation
We prepared cross-sectional thin specimens for HAADF-STEM characterisation using a focused ion beam with a Ga + ion source on a FEI Helios G4 UX instrument. Before the milling process, a ∼5 nm-thick Au layer was deposited on the film surface to protect and enhance the conductivity of sample during milling. The lift-out lamellae were thinned to ∼100 nm at 30 kV with a current decreasing from 0.75 nA to 90 pA, followed by final polishing at 2 kV and 17 pA. HAADF-STEM and NBED images were obtained using a Cs-corrected FEI Titan G2 80-200 equipped with a Super-X EDS. The HAADF-STEM images were collected using a convergence semi-angle of 18 mrad and an inner-collection semi-angle of 55 mrad. EDS mapping was performed using a Bruker Esprit analysis system with automatic drift correction during the collection process, which augments the reliability of the compsitional analysis. Integrated line profiles were conducted across the heterostructure in EDS maps to enhance the signal-to-noise ratio. The Cliff-Lorimer analysis method was applied to quantify the EDS line-scan results.