Introduction

Topological materials have attracted intensive attentions due to their robust topologically protected states, many exotic properties and promising applications for quantum computing and spintronics1,2. According to the dimensionality of electronic bands touching, the topological states of materials can be classified into topological insulators3,4, Dirac semimetals5,6 and Weyl semimetals7,8, etc. Weyl semimetal has the feature of Weyl fermion with the presence of the chiral node (i.e. Weyl node) and the Fermi arc surface states connecting the Weyl-node pair with opposite chirality9. The Weyl node is a linearly crossing point of two non-degenerate bands which requires breaking inversion symmetry or time reversal symmetry. In order for developing electronic device, it is essential for effective electrical manipulation of the nontrivial topologic states e.g. Weyl nodes. Magnetic Weyl semimetal is considered as an ideal material candidate since the time reversal symmetry is breaking and the location and enegy of Weyl nodes in the Brillouin zone depend on the magnetization direction2. Mn3Sn is a typical Weyl semimetal and non-collinear antiferromagnet (AFM)10,11,12,13. As shown in Fig. 1a, the spin structure of Mn3Sn consists of two kagome planes with opposite chirality. This hexagonal spin texture can be considered as a ferroic ordering of a cluster magnetic octupole M and it breaks time reversal symmetry macroscopically. AFMs have negligible stray field and ultra-fast magnetic dynamics, which helps to overcome the integrability and speed bottlenecks of traditional spintronic devices14,15,16,17. Furthermore, all-AFM-based magnetic tunnel junctions (MTJ) with a sizable tunneling magnetoresistance (TMR) ratio have recently been demonstrated18,19.

Fig. 1: Schematic of OT-driven magnetization switching in Mn3Sn.
figure 1

a Spin structure of Mn3Sn. The large blue and red circles (small black and silver circles) represent Mn (Sn) atoms. In two kagome planes with different chirality, blue and red arrows indicate the magnetic moments of Mn atoms in different layers. The green arrow indicates the M direction of the formed cluster magnetic octupole. b Schematic of ST-driven magnetization switching in Mn3Sn. Current in spin source layer can generate spin angular momentum S+(-) which is injected into Mn3Sn layer and exerts a torque on the kagome planes of Mn3Sn. c Schematic of OT-driven magnetization switching. Before being exerted on Mn3Sn, orbital angular momentum L+(-) induced by current in OT source layer needs to be converted to S+(-). d Stack structure of Mn3Sn/Cu/CuOx film and electrical measurement setup. e Longitudinal resistance of Mn3Sn/Cu device as a function of oxidation time. Inset illustrates RAHE versus applied magnetic field. f Current-induced magnetization switching loops in as-deposited Mn3Sn/Cu device and in Mn3Sn/Cu/CuOx device.

Full size image

To date, researchers have demonstrated that spin torques (ST) can manipulate the Weyl nodes in Mn3Sn or Co2MnGa manifested with the change in anormolous Hall effect (AHE)20,21,22,23,24, using a similar protocol as for heavy metal/ferromagnets (HM/FM)25,26,27,28,29 where the spin current generated in heavy metal by spin Hall effect (SHE) is injected into FM and a torque is exerted on FM. As shown in Fig. 1b, the generated spin current induced by current along <0001> direction can be directly exerted on the kagome planes of Mn3Sn and induce the magnetization switching. The switching efficiency largely depends on the charge-current-to-spin-current conversion efficiency, i.e., the spin Hall angle (SHA), in the adjacent spin current source layer. To improve the efficiency and reduce the switching current density Jc, except for exploring novel materials with high SHA, it is also desirable to search alternative mechanisms to realize current-induced switching of topological states with high energy efficiency.

In this work, we propose to utilize orbital torques (OT) which originate from orbital Hall effect (OHE) or orbital Rashba-Edelstein effect (OREE) to manipulate the magnetic order in topological AFM. The basic schematic is shown in Fig. 1c. An applied current along <0001> direction can induce orbital current in OT source layer30,31,32. Before the generated orbital current can be exerted on the magnetization of Mn3Sn, it need to be converted to spin current by an additional spin-orbit coupling (SOC)33. Compared with the limited SHA in spin current source material, it has been demonstrated that orbit current source materials possess a much higher orbit current generation efficiency34,35. Therefore, the critical switching current density Jc is expected to be effectively reduced. Herein, we successfully demonstrate OT-driven magnetization switching in Mn3Sn and prove that the orbit-current-to-spin-current (L-S) conversion can be done either by Mn3Sn itself or by inserting a heavy metal with strong SOC. we have achieved Jc as low as ~1 × 1010 A/m2, which is more than one order of magnitude lower than the common Jc in SHE-driven framework. Furthermore, we show that the stable memristor-like switching characteristics offers Mn3Sn with excellent plasticity to mimic an artificial synapse with linear potentiation and depression processes, which is beneficial for constructing neural networks with high accuracy.

Results

OT-driven magnetization switching in topological Mn3Sn

We deposited 40-nm-thick Mn3Sn film on thermally oxidized silicon substrates by magnetron sputtering (see details in Methods). By energy dispersive spectrometry (EDS) mapping, the atomic percentage of MnSn alloy is determined to be around Mn(3.05-3.1)Sn. The x-ray diffraction (XRD) θ−2θ scans results of the deposited Mn3Sn is shown Supplementary Note 1. Compared with pure Si substrate, a clear Mn3Sn (0002) crystal peak is observed in the film sample. SQUID measurement indicates that our film exhibits a tiny out-of-plane magnetization (see Supplementary Note 1). This tiny out-of-plane hysteresis loop suggests there exist crystalline grains with its kagome plane in the film normal since the spin canting is in the (0001) kagome plane. We further carried out magneto-transport and anormolous Nerst effect (ANE) measurements to confirm the Weyl semimetal of our deposited Mn3Sn films. Planar Hall effect (PHE) and in-plane angular magnetoresistance (AMR) are shown in Supplementary Figure 1c. The PHE and AMR follow the functions \({R}_{{xy}}=-\Delta R\sin \theta \cos \theta\) and \({R}_{{xx}}={R}_{\perp }-\Delta R{\cos }^{2}\theta\), respectively, where \(\Delta R={R}_{\perp }-{R}_{\parallel }\), and \({R}_{\perp }\) and \({R}_{\parallel }\) are the resistances when the magnetic field directions are perpendicular and parallel to the charge current direction, respectively. These are consistent with the feature of the chiral anomaly induced PHE and AMR in Weyl semimetal9,20. Comparing to the effect of magnetization, ANE in Weyl semimetal is much enhanced due to the Weyl nodes around Fermi level11. The large ANE and small magnetization further confirm that our Mn3Sn films are the Weyl semimetal (Supplementary Figure 1b and c).

It has been widely demonstrated that OT can be observed in the naturally oxidized Cu36,37. Thus, we firsly choose oxidized Cu to verify the impact of OT on the magnetization switching of Mn3Sn antiferromagnet. As shown in inset of Fig. 1d, we firstly deposited a Cu layer on the Mn3Sn layer and then follow the method in previous works to naturally oxidize Cu at atmosphere for different time37. The films were fabricated into Hall bar device of 5 μm width to implement magneto-transport measurement. The schematic setup of the measurement is illustrated in Fig. 1d. The longitudinal resistance shows a continuous increase with oxidation time, which suggests the gradual oxidation of Cu layer with the time. We then measure the AHE resistance RAHE (inset in Fig. 1e) as a function of the out-of-plane magnetic field to estimate the switchable magnetic domains which corresponds to the crystalline grains with Kogame plane in the film normal. In the absence of magnetic field, there exist two stable magnetic states which correspond to the magnetic octupole M along ± z directions, respectively.

We then carry out current-induced switching experiments (Fig. 1f). Since Cu is a light metal with negligible SOC, current-induced magnetization switching is absent in the as-deposited Mn3Sn/Cu film. As a comparison, a deterministic magnetization switching loop, corresponding to a switching ratio of ~ 25%, is well achieved in the device after natural oxidatization process of the Cu. We are also aware that several works reported current-induced switching in Mn3Sn single layer with specific crystal configuration38,39. However, we didn’t observe any switching phenomenon in our deposited Mn3Sn single layer (Supplementary Note 2), verifying that the magnetization switching driving force indeed comes from the achieved Cu/CuOx layer. Additionally, in Supplementary Note 3, we demonstrated that only when the applied FM like Ni exhibits a high SOC, one can observe a sizable effective SHA in FM/Cu/CuOx heterostructure40. From one side, it verifies that the main switching driving force from Cu/CuOx is OTs instead of possible STs. From the other side, it emphasizes again that the orbital current originated from CuOx must complete the L-S conversion process shown in Fig. 1c to manipulate the magnetic dynamics in FM. Our deterministic switching results in Mn3Sn/Cu/CuOx device directly prove that Mn3Sn itself can complete the L-S conversion process like what Ni does and the spin current converted from orbit current is then to switch the magnetization of Mn3Sn layer. More experimental evidences related to SOC and L-S conversion in Mn3Sn will be presented in the next section.

Note that, for practical use, it is important to quantitatively control the orbital Hall angle in the device. In such case, metallic OT sources have an application edge over naturally oxidized Cu. We thus employ an heterostructure composed of Mn3Sn (40 nm)/Pt(tPt)/Mn(tMn) trilayer (Fig. 2a) to further investigate the detailed OT-based manipulation characteristics of topological magnetization. Mn is theoretically predicted to possess a large orbital Hall angle (~18)35. The inserted Pt layer serve as an additional L-S conversion layer which helps to gain more spin torques33,41. In a device where tPt = 2 nm and tMn = 10 nm, deterministic OT-driven magnetization switching in Mn3Sn is achieved as well (see detailed switching loops in Supplementary Note 4). An external magnetic field Hex is required to break the in-plane symmetry during the switching process. When reversing the direction of Hex, the switching polarity also changes. This switching characteristic is very similar to the SHE-induced switching protocol for FMs. Moreover, we have observed a Hex-dependent switching ratio trend in the sample. The switching ratio is defined as ΔRc/ΔRH, where ΔRc and ΔRH are current-induced and field-induced change of RAHE. As shown in Fig. 2b, as the absolute value of Hex increases, the switching ratio will first increase and then saturate when exceeds 2 kOe. The saturated switching ratio is around 27%, which is comparable with other reported value20,42,43. This relatively small switching ratio can be explained by the fact that in-plane torques only allow the kagome planes to rotate between two energy minimum states with θ = ±π/6 (see Fig. 2c) according to the symmetry analysis20. Note that all the following switching experiments in this work were carries out under Hex = 2 kOe, unless specified. Additionally, in Supplementary Note 5, we show that ANE signal switches simultaneously with the anomalous Hall signal, suggesting that non-trivial topology in Mn3Sn can also be manipulated by the orbital torque.

Fig. 2: OT-driven magnetization switching in Mn3Sn/Pt(2 nm)/Mn(10 nm).
figure 2

a Schematic of the Mn3Sn/Pt/Mn heterostructure. b OT-driven switching ratio as a function of Hex. The ratio saturates when Hex exceeds 2 kOe. Red (purple) region indicates the saturated switching ratio area in the positive (negative) magnetic field range. c Schematic of two stable magnetic states in Mn3Sn during OT-driven switching dynamics. d RAHE-J switching loops with different applied pulse width. e Minor switching loops of the sample by limiting the maximum current value Jmax in negative range, as indicated by the numbers in the figure.

Full size image

An important difference between switching in FM and switching in Mn3Sn is the impact of applied current pulse width. In FM system, a thermally activated switching model, in which Jc decreases exponentially with the increasing pulse width, is widely accepted44. In Fig. 2d, we plot RAHE as a function of the applied charge current density J in Pt/Mn bilayer with different pulse width. Clearly, Jc in Mn3Sn is almost insensitive to the current pulse width which varies from 10 μs to 500 μs (see Fig. 2d). This insensitivity indicates that our deposited Mn3Sn possesses a good thermal stability, which is beneficial for the device scalability. We also notice that the achieved RAHE-J switching loops are quite ‘tilted’, i.e., there exists a series of intermediate states. As shown in Fig. 2e, a series of minor switching loops can be achieved by limiting the maximum value at negative current range. In Supplementary Note 6, we show that controlling the applied out-of-plane magnetic field can also achieve similar minor loops. The stable existence of these minor loops reveals that the intermediate states are non-volatile and can be recovered to the same initial state by applying a current density J ~ 9 × 1010 A/m2. This switching characteristic indicates that the Mn3Sn device can possibly memorize the past electrical current pulse and be adapted as a memristor. We will further investigate the potential application of this memristor-like behavior in neuromorphic computing in the last section.

OT source layer dependence of switching efficiency

To further verify that the observed switching behaviors are dominant by OTs in Mn3Sn/Pt/Mn and to optimize the switching performance, we implemented current-induced switching experiments in samples with different Pt and Mn thicknesses. We first investigate the impact of tPt by fixing tMn to 10 nm and varying tPt from 0.5 nm to 6 nm. In all the samples, deterministic switching is observed, while the switching polarity remains the same (see Supplementary Note 4). Figure 3a plots Jc as a function of Pt thickness tPt. This trend can be separated into 3 stages: 1) when tPt ≥ 4 nm, Jc keeps at a stable plateau; 2) when 1 nm ≤ tPt < 4 nm, Jc decreases with t decreasing tPt; 3) when tPt < 1 nm, Jc starts to slightly increase with decreasing tPt. It is known that the effective SHA of Pt will first increase and then saturate when tPt increases from 0 nm to 4-5 nm45. The change of Jc with tPt can be understood as follows. At stage 1, STs from Pt dominate the switching and OTs barely participates in the process. At stage 2 and 3, OTs gradually dominate the switching process, since the observed Jc-tPt trend in this region is opposite to the Jc-tPt trend in conventional ST-dominant system (see detailed switching results of Mn3Sn/Pt in Supplementary Note 2).

Fig. 3: OT source layer dependence of the switching efficiency.
figure 3

a Jc as a function of Pt thickness tPt with a fixed Mn thickness (tMn = 10 nm). When tPt ≤ 3 nm, the switching dynamics are dominated by OT (red region). When tPt > 3 nm, ST (blue region) dominates the switching dynamics. The error bars in (a) and (c) are obtained from multiple switching loops for each sample. b Effective SHA in Co/Pt/Mn as a function of tPt while tMn remains at 10 nm. Inset shows the film stack. The error bars in (b) and (d) are the standard deviation of SHA at different frequency. c Jc as a function of tMn with a fixed tPt (2 nm). A monotonic decreasing trend can be achieved. d Effective SHA as a function of tMn while tPt remains at 2 nm. e First harmonic Hall resistance Rω and the fitting curve as a function of β in Mn3Sn/Pt(2 nm)/Mn(20 nm) device. The applied magnetic field is 6 T, while the applied current I is 1 mA. Inset shows the measurement setup. f Second harmonic Hall resistance R2ω and the fitting curve as a function of β.

Full size image

The minimum Jc appears at around tPt = 1 nm, revealing that the L-S conversion efficiency in Mn3Sn/Pt/Mn system maximizes at this point. To better quantify the switching efficiency, we measured the effective SHA in Co/Pt(tPt)/Mn(10 nm) heterostructure by spin-torque ferromagnetic resonance (ST-FMR) technique (see details in Supplementary Note 7). As shown in Fig. 3b, when tPt increases from 1 nm to 6 nm, the effective SHA will first increase and then decrease. The highest SHA ( ~ 0.4) appears when tPt is around 2-3 nm, which is comparable with other reported values33,34,41. We also notice that this optimal SHA point is shifted from the optimal Jc point in Mn3Sn/Pt/Mn system. A key parameter related to the SOC strength is the spin diffusion length λsf. The shorter is λsf, the stronger is the SOC. The experimentally determined λsf in Mn3Sn is ~0.75 nm46, which is one order of magnitude shorter than λsf in Co (7-12 nm)47. It confirms that the Mn3Sn has larger SOC than Co, which can explain why we observed deterministic switching in Mn3Sn/Cu/CuOx device but we failed to extract spin torque signal in Co/Cu/CuOx device (Supplementary Note 3). The weaker SOC in Co than that in Mn3Sn will lead to thicker Pt layer for the optimal effective SHA, which is the reason of the inconsistency between Co and Mn3Sn based samples.

We then investigate the impact of tMn by fixing tPt to 2 nm and varying tMn from 3 to 20 nm. As shown in Fig. 3c, a monotonic decreasing Jc-tMn trend can be achieved. When tMn = 20 nm, Jc is reduced to ~1 × 1010 A/m2, which is more than one order of magnitude lower than Jc in Mn3Sn/Pt bilayer20. This trend is consistent with the tMn-dependent SHA trend in Co/Pt/Mn system (see Fig. 3d). When tMn increases from 10 nm to 20 nm, the effective SHA monotonically increases. At tMn = 20 nm, the effective SHA is determined to be around 0.64. We consider this unsaturated effective SHA within a large tMn range as another important characteristic of OT. According to the drift-diffusion equation, the orbital Hall angle of Mn θMn, which is defined as the charge-current-to-orbital-current conversion efficiency, can be described by θMn = σMn[1-sech(tMn/λMn)], where σMn and λMn are the orbital Hall conductivity and orbital diffusion length of Mn, respectively. λMn (~ 11 nm) is theoretically expected to be much longer than the typical spin diffusion length of conventional heavy metal (1-2 nm for Pt)34. As a result, θMn should have a much longer saturation length (> 20 nm in our work) than the saturation length of effective SHA in Pt (typically 5 nm). Given a fixed L-S conversion efficiency, a larger θMn will certainly lead to a larger effective SHA in the system as well as a lower Jc. Note that, we here consider the average current density in Pt/Mn bilayer. In Supplementary Note 8, we show that all the trends and conclusions are still solid regarding the seperated current density distributions in Pt layer and Mn layer. We also show that our OT-driven switching scheme not only reduces Jc, but also reduces the switching power consumption by one order of magnitude, which leads to low-power benefit in realistic application (Supplementary Note 9).

To better quantify the actual effective spin Hall angle in our Mn3Sn/Pt/Mn device, we also implemented harmonics measurement (see measurement setup in inset of Fig. 3e)48. When we apply an ac current I along x axis and rotate the external magnetic field in xz plane, the octupole moment Δφoct will rotate coherently and result in the change of first harmonic signal Rω in xy direction (Fig. 3e). The oscillation of Δφoct will also leads to a second harmonic signal R2ω in the form of \({\left.\frac{d{R}_{\omega }}{2d\Delta {{{{{{\rm{\varphi }}}}}}}_{{oct}}}\right|}_{I=0}\Delta {{{{{{\rm{\varphi }}}}}}}_{{oct}}(I)\) (Fig. 3f). The current-induced octupole oscillation \(\Delta {\varphi }_{{oct}}\left(I\right)\) can be calculated using the torque balance equation (see calculation details in Supplementary Note 10) The fitting results allow us to obtain the damping-like effective field HDL. We can then calculate the effective SHA by \({SHA}=\frac{2e{\mu }_{0}(3{M}_{0})t{H}_{{DL}}}{\hslash J}\), where \({M}_{0}\), t, ℏ and JSOT are the magnetization of a sublattice moment, the Mn3Sn thickness, the reduced Planck constant and the average current density in the source layer, respectively. In our work, the calculated effective SHA in Mn3Sn/Pt(5 nm) is only ~ 0.026. As a comparison, the effective SHA in Mn3Sn/Pt(2 nm)/Mn(20 nm) is determined to be ~ 0.32, which is more than one order of magnitude higher. This large SHA difference also well corresponds to the Jc difference and demonstrate again that OT-driven Mn3Sn switching is of high efficiency.

Neuromorphic computing based on OT-based manipulation of Mn3Sn

Artificial synapses are considered as an ideal hardware to implement neuromorphic computing49,50,51,52. Recent works suggest the current-induced magnetization switching process in ferro- and ferri-magnetic materials can mimic the long-term depression (LTD) and the long-term potentiation (LTP) functions of synapses, following a general domain nucleation theory53,54. However, the linearity of achieved LTD (LTP) processes, which is considered as a fundamental parameter for constructing high-accuracy artificial neural network (ANN), has arrived at a ceiling, because of the limited magnetic domain size. In such case, Mn3Sn is expected to be a better material platform, since the AFM nature of Mn3Sn can reduce the magnetic dipole effect and thus allow the existence of more tiny magnetic domains in the crossbar area than ferromagnet does. Moreover, the random magnetic domain switching phenomena brought by thermal fluctuation is also expected to be suppressed, given the low OT-driven critical switching current density as well as the good thermal stability of Mn3Sn.

In Fig. 2e, we have already shown that the OT-driven switching process of Mn3Sn exhibits a memristor-like behavior. Here, given the normalized RAHE detected during the switching can be defined as the weight (G) of the artificial synapse, we demonstrate the realization of LTD and LTP functions in a Mn3Sn/Pt(2 nm)/Mn(10 nm) device. As shown in Fig. 4a, to achieve such processes, 120 negative (positive) current pulses of fixed amplitude are applied. The current pulse amplitude is ~4.5 × 1010 A/m2, which corresponds to the very beginning switching point of the device. Interestingly, we found the achieved LTD (LTP) processes show very high linearity. To better quantify the linearity performance, the nonlinearity (NL) of weight update is defined as \({NL}=\frac{\max \left|{G}_{P}\left(n\right)-{G}_{D}\left(121-n\right)\right|}{({G}_{\max }-{G}_{\min })}\) for n = 1 to 120, where GP(n) and GD(n) are the normalized RAHE values after the nth potentiation pulse and nth depression pulse. Gmax and Gmin represent the maximum GP(n) after 120 potentiation pulses and minimum GD(n) at initial state. Here, a very small NL value (∼0.166) is determined, indicating that our Mn3Sn-based artificial synapse possesses a similar LTD (LTP) process to an ideal device. As a comparison, if we apply the ST-driven Mn3Sn switching scheme, i.e., Mn3Sn/Pt device, to realize the LTD and LTP process, the obtained NL value (~0.686) is much larger, indicating a worse linearity. See Supplementary Note 11 for detailed discussion.

Fig. 4: ANN system with Mn3Sn-based artificial synapse.
figure 4

a LTD and LTP process with high linearity in Mn3Sn/Pt(2 nm)/Mn(10 nm) device. The calculated NL is as low as 0.166, which indicates a good linearity. b Schematic of the constructed ANN with 100 × 100 memory cells for image recognition task. c Evolution of the images in the learning processes with the experimental and ideal devices. The image is taken from Yale Face Database B55,56. d Image accuracy rates as a function of learning epochs in constructed ANN using three kinds of devices.

Full size image

To better evaluate the performance of the proposed synapse, an ANN with a modelled 100 × 100 Mn3Sn-based memory array was then constructed to implement pattern recognition task (see Fig. 4b). The subsequent learning processes followed the synaptic weight change processes shown in Fig. 4a. Each memory cell serves as a synapse to connect pre- and post-neurons, and the synaptic weight of each cell was represented by the gray level of each pixel. The 100 × 100 pixels of initial input image taken from the Yale Face Database B is employed for the pattern recognition task55,56, and the gray variation of each pixel is real-time stored in each cell with the increasing learning epochs. Figure 4c compares the image evolution with various numbers of learning epochs during the learning process for the experimental and ideal devices. For the quantitative analysis of learning efficiency, the learning accuracy at every 5 epochs, which is defined as the difference between the original image and learned image, can be obtained (see details in Supplementary Note 6). As shown in Fig. 4d, the learning accuracy rate of the OT-driven Mn3Sn device is 97.5%, which is only slightly lower than that of the ideal device (∼99.5%). Meanwhile, the lowest learning accuracy is achieved in the ST-driven Mn3Sn device, due to the worst linearity in its LTD and LTP process. The above simulation results strongly suggest the high application potential of our OT-driven Mn3Sn spintronic device in neuromorphic computing. We note that recent studies suggest the presence of memristive behavior in conventional collinear antiferromagnets (AFMs) when subjected to both electrical and optical manipulation57,58,59. On one hand, since the time reversal symmetry (TRS) is conserved, the Néel order in collinear AFMs is usually switched by 90 degrees and remains hard to be detected. In terms of reliable readout, noncolinear AFMs with broken TRS have an advantage over colinear AFMs. On the other hand, we are aware that the ultra-fast switching of noncolinear AFMs at picosecond timescale hasn’t been reported experimentally, despite of theoretical predictions. We expect that the development of effective methods to achieve picosecond manipulation of the Néel order in noncollinear AFMs could further enhance the performance of the constructed ANN.

Discussion

In summary, we have demonstrated that OTs can serve as an effective electrical method to manipulate the topological magnetization of antiferromagnets. We prove that Mn3Sn can directly convert the orbital current from the OT sources, e.g., metals (Mn) and oxides (CuOx), to spin current. We also show that an inserted Pt layer between Mn3Sn and OT source can enhance the L-S conversion efficiency. By adjusting the thickness of Pt and Mn, the critical switching current density can be reduced to as low as ~1 × 1010 A/m2. In addition, we show that the OT-driven switching process in Mn3Sn can mimic the LTD and LTP process with high linearity in an artificial synapse, which can be further utilized to construct ANN system with high image recognition accuracy. Finally, given TMR has been reported in AFM-MTJ, it is possible to incorporate the presented OT-driven Mn3Sn switching scheme in an OT-based AFM MTJ devices. Our finding goes beyond the conventional paradigm of using spin current to manipulate AFM’s order, and offers an alternative to integrate topological AFM in future diverse spintronic devices.

Methods

Sample growth and device fabrication

Mn3Sn(40)/Pt(0-6)/Mn(0-20)/AlN(5) and Mn3Sn(40)/Cu(10) stacks (thickness in nm) are deposited on thermally oxidized silicon substrates by DC and RF magnetic sputtering (AJA) under a base pressure lower than 3 × 10−8 Torr. AlN is an insulating capping layer. Mn3Sn are deposited by co-sputtering of Mn2.5Sn and Mn targets at room temperature, followed by annealing in situ at 500 °C for 1 hour. After cooling down to room temperature, the following Pt/Mn/AlN or Cu layers are deposited. For the oxidation process, Mn3Sn/Cu films were exposed to dry atmosphere for a set time. Then, AlN is deposited to stop Cu from being further oxidized. The device is fabricated via an Ultraviolet Maskless Lithography machine (TuoTuo Technology). The etching is done by standard ion milling technique.

Electrical measurement

For transverse and longitudinal signal measurements, an ac current of 317.3 Hz was applied along the x axis, while SR830 and Zurich lock-in amplifiers are used to detect the voltage. For current-induced switching and neuromorphic functions, electrical current pulses (width from 10 to 500 μs) were applied. After each pulse, we wait for 5 s to avoid the Joule heating and use a small ac current to read out the AHE voltage. For ST-FMR measurement, a Rohde & Schwarz SMB 100 A signal generator was used to provide the modulated microwave. The rectifying voltage was collected using a lock-in amplifier.