The unprecedented success of silicon-CMOS technology has been primarily driven by transistor scaling. The early era of Dennard scaling1,2 had two important consequences for modern computing systems; speed performance has been improved by scaling of physical dimensions while power density has been kept practically constant by scaling of voltages. However, since around 2005, the voltage scaling has become a challenging task3 because further reduction of the supply voltage (requiring the threshold voltage to be simultaneously decreased to maintain the capability to drive high current) was leading to an exponential increase in the leakage current. This is due to the fundamental limit of subthreshold swing (SS) greater than 60 mV/decade at room temperature, which arises from the Boltzmann statistics that govern the thermionic operation of conventional MOSFETs4. Although tunnel FETs5,6,7, negative capacitance FETs8,9,10, and electrostrictive FETs11,12 have recently emerged as novel device concepts for steep SS, they still require much work to improve the device performance and/or answer fundamental questions on scalability, reliability, and viability. On the other hand, scaling of physical dimensions still continued until today, towards the technology nodes of 5 nm and beyond13,14. This led to an inevitable increase in power density in high-performance microprocessors, thus requiring complex power management techniques15,16. In view of this, spintronics, which utilizes the quantum mechanical property of elementary particles, called spin, has the potential to become an innovative pathway beyond transistor scaling to satisfying the speed and energy-efficiency needs of the emerging computing paradigm (e.g., neuromorphic17,18 or quantum19,20 computing).

Spintronics is an emerging form of electronics that harnesses the intrinsic angular momentum (spin) of electrons, leading researchers to treat them as tiny bar magnets with strong magnetic moment (Bohr magneton, μB ~ 10−23 Am2). Immediately after the discovery of giant magnetoresistance (GMR) in 1988 (the GMR ratio of 10–50%, as compared with the anisotropic MR effect of at most a few percent at that time)21,22, which has been the foundation for constructing the hard-drive read head, research on nanomagnetism and spintronics has made remarkable strides over the last few decades. In 1991, the oscillatory behavior of the interlayer exchange coupling was formulated in terms of the Fermi surface of the spacer material23,24, leading a way to constructing a synthetic magnetic free layer in the modern magnetic tunnel junction (MTJ) device25,26. The concept of “spin valve” ignited numerous research efforts exploring the ferromagnetic sandwiches spaced by a nonmagnetic metallic layer27,28; logical “1” or the high-resistance state is represented by the antiparallel configuration of two ferromagnetic layers while logical “0” or the low-resistance state is dictated by the parallel configuration. This is an important milestone in spintronics because spin valve currently serves as the most fundamental building block for advanced spintronic devices. In 1995, insertion of a thin insulator as the spacer layer (i.e., as the tunnel barrier) resulted in a much higher read-out signal through the effect of TMR (tunneling magnetoresistance)29, which continued to be further optimized to reach 604% at room temperature in 200830. Combined with the write mechanism of spin-transfer-torque (STT), theoretically predicted in 199631,32 and observed in the experiments of switching the magnetization direction33 and inducing domain-wall motion34 or high-frequency oscillation35, both the spin valve and MTJ structures (especially, MTJs employing the crystalline MgO tunnel barrier36,37) have gained great attention in the field. The recent discovery of the perpendicular magnetic anisotropy (PMA) in various heterostructure thin films and interfaces offered a new opportunity for low-power data storage applications38,39,40. The reviving interest in the Hall effect41 will link the spin Hall effect (SHE) with measurable quantities such as spin accumulation, aiming at driving magnetization dynamics more efficiently for the development of energy-efficient spintronic devices.

The ability of spintronics to perform previously impossible experiments and test fundamental theories of magnetism has recently been made possible through the use of 2D materials. Layered materials have, indeed, existed for eons (i.e., studied scientifically for more than 150 years42). However, only recently have researchers begun to realize the true potential of these layered materials for advanced technological applications. Each layered material, when thinned down to its physical limits, exhibits novel properties different from its bulk counterpart. Therefore, these materials are specifically referred to as “2D materials”. The most widely studied 2D material is graphene43, which is an atomically-thin sheet of carbon atoms with a bonding length of 1.42 Å. Pristine graphene behaves as a zero-bandgap semimetal, constituting an important class of carbon allotropes exhibiting unique 2D hexagonal structures. Readers interested in recent advances and a broad range of commercial applications of graphene will find the 2019 review article by Kong et al.44 an excellent resource. A wide spectrum of other 2D inorganic materials is also experiencing numerous research efforts owing to their versatile crystal structures and physical properties, including insulators (e.g., hexagonal boron nitride, h-BN)45, semiconductors (e.g., transition metal dichalcogenides such as MoS2)46,47, superconductors (e.g., NbSe2)48, and even magnets (e.g., CrI3)49.

This review discusses 2D materials in the context of spintronic device technologies, focusing on their progress, prospects, and challenges. We examine how 2D materials can be interfaced with other classes of materials, including ferromagnets (FMs), to enable, develop, and advance spintronic devices. Spin-dependent properties of 2D materials are also highlighted in terms of spin-orbit and spin-valley couplings. Major challenges and future perspectives for the development of spintronic devices using 2D materials are finally discussed.

Graphene for spin logic

Among various device concepts for beyond-CMOS logic switches50, spin-based, non-volatile digital logic51,52,53,54,55,56,57,58 (Fig. 1a) has attained much attention because it offers the functionality at lower power consumption by building logic units with magnetic elements. In particular, the proposal of all-spin logic (ASL) switch51 has been considered to be quite attractive for a future computing paradigm due to its potential for extremely low-power consumption59,60. This is because, in the ASL, nonlocal spin currents61,62,63 are used to selectively route communication signals between the input and output magnets without spin-to-charge conversion (Fig. 1b). However, a major challenge has been to develop a suitable spin transport channel material that simultaneously achieves a high spin injection efficiency (from FM to the channel material) and a long-distance spin propagation (within the channel material). Graphene, a hexagonal 2D arrangement of sp2-bonded carbon atoms, may be uniquely suited for this purpose due to its (1) relatively long spin diffusion length and lifetime (theoretically predicted, and experimentally being examined) and (2) ability to be readily interfaced with other classes of materials, including FMs.

Fig. 1: Spin logic devices.
figure 1

a Reconfigurable spintronic logic gate52, showing a NAND operation between X and Y. A and B are chosen to be “1” for NAND, but one may adjust the magnetization alignments of A and B to set different logic operations. The output is given by a transient current response, IM(t), whose amplitude is proportional to the spin accumulation in the semiconductor. A transient current is allowed to pass through the probing contact (M) during the magnetization rotation, where the potential level inside M is perturbed and the capacitor (CM) tries to restore the new steady-state condition (IM = 0). In order to write the magnetization directions of the contacts in cascaded logic gates, IM(t) needs to be converted to a binary voltage signal. Vdd is the supply voltage. b All-spin logic (ASL) switch51, where nonlocal pure spin currents transport information between nanomagnets (i.e., digital spin capacitors). The state of the output magnet is determined by two spin torques with the information provided from the input magnet: one relatively large torque is exerted by the application of Vsupply and the other small torque is exerted by the application of Vbias (for “receive” and “send” respectively, in the cascaded logic gates). 1: magnetic free layer, 2: isolation layer (electrostatic barriers or insulation layers), 3: tunneling layer (oxides or Schottky barriers), 4: spin-coherent channel, 5: contact, 6: spacer layer, and 7: magnetic fixed layer. Image of a reprinted by permission from Springer Nature: ref. 52, Copyright 2007. Image of b reprinted by permission from Springer Nature: ref. 51, Copyright 2010.

Spin injection, transport, and relaxation in graphene

Graphene exhibits properties attractive for spintronic device applications64,65. It has intrinsically low spin-orbit coupling (SOC); at the Γ point, the spin-orbit splitting of the σ band is calculated to be as low as 8.8 meV66,67. This can lead to spin coherence maintained over the long distance. Graphene also features ultra-high carrier mobility of reaching up to more than 200,000 cm2 V−1s−1 in a suspended device structure68, which can enhance the diffusive transport of spin. Advances in spin injection into graphene, experimental determination of spin transport properties of graphene, and understanding of the spin relaxation mechanism in graphene will ultimately contribute to the development of the spin logic device adopting graphene as a spin transport channel.

Nonlocal spin valve (NLSV)

The NLSV63,69 is the most commonly used device geometry for generating pure spin currents. When the spin-polarized current is injected from a FM contact into a channel material, the accumulated spin in the vicinity of the FM/channel interface can diffuse laterally, thus forming a pure spin current. This nonlocal spin current constitutes a key part of the ASL switch, enabling the input logic bit to control the state of the corresponding output through the spin-torque mechanism70. Moreover, the spin-dependent properties of graphene have been quantitatively understood by measuring the NLSV built upon graphene71, consisting of two FM (one as a spin injector and the other as a spin detector) and two corresponding nonmagnetic (NM) metallic contact electrodes integrated on top of the graphene spin transport channel (Fig. 2a). In this graphene-based NLSV structure, sweeping the in-plane magnetic field (Hy) allows both parallel and antiparallel configurations of two FM electrodes with different shape anisotropy, thus resulting in the nonlocal spin signal, ΔRNL = ΔVNL/I (Fig. 2b). ΔVNL is the nonlocal voltage difference between parallel and antiparallel states of two FMs measured at the FM detector side (ΔVNL = VNLP – VNLAP) and I is the injection current through the FM injector. This four-terminal measurement geometry is called nonlocal because the voltage is measured “outside” the (charge) current loop. One key advantage of this NLSV geometry is that background noise signals from charge currents are greatly reduced due to the absence of net charge flow. Importantly, the nonlocal spin signal will only be dependent on the relative orientation of the FM detector’s spin with respect to those accumulated at the FM injector. Thus, it is possible that ΔRNL is measured by varying the injector and detector separation (L) because the magnitude of spin accumulation will decay as it diffuses away from the injection point, with a characteristic length known as spin diffusion length (LD). This is described by the equation below (Eq. 1)69.

$$\Delta R_{NL} = \frac{{\eta ^2L_D}}{{2W\sigma }}exp\left( { - L/L_D} \right)$$

In this equation, σ is the conductivity, W is the width of the graphene channel, and L is the contact spacing. η is the spin injection efficiency, often represented by the spin polarization of the injection current \(( {\frac{{\mathrm{I}}_{\mathrm{majority}}-{\mathrm{I}}_{\mathrm{minority}}}{{\mathrm{I}}_{\mathrm{majority}} + {\mathrm{I}}_{\mathrm{minority}}}})\), where majority (minority) means the spin direction parallel (antiparallel) to that of the FM injector. Fundamentally, η is directly related to the spin polarization of the FM contact electrode \(({\frac{{\mathrm{DOS}}_{\mathrm{majority}}-{\mathrm{DOS}}_{\mathrm{minority}}}{{\mathrm{DOS}}_{\mathrm{majority}} + {\mathrm{DOS}}_{\mathrm{minority}}}})\), where DOS is the density of states at the Fermi level, modulated by any presence of spin-flip scattering at the FM/channel interface. It is noted that above definitions, including Eq. 1 and spin polarization of FM electrodes, are only valid for the case of tunnel barrier contacts.

Fig. 2: Graphene nonlocal spin valve.
figure 2

a Schematics of the graphene nonlocal spin valve (NLSV), showing both the side and top view images with graphene as a spin transport channel. FM and NM represent ferromagnetic and nonmagnetic metals, respectively. The spin transport properties of graphene can be determined by measuring the difference in nonlocal resistances between parallel and antiparallel states of two FMs (ΔRNL) with varying the injector and detector separation (L), or measuring the Hanle curve by applying an out-of-plane (Hz) magnetic field to induce a spin precession. b The nonlocal resistance measurement with the graphene NLSV that has the FM contact separation distance (L) of 2.1 µm and the monolayer graphene channel width (W) of 2.2 µm82. Co (Cobalt) is used in this work to represent the four contact electrodes (E1 through E4) with the MgO/TiO2 bilayer tunnel barrier inserted at the Co and graphene interface. The black (red) curve shows the nonlocal resistance as the magnetic field (H) is swept in the +y (−y) direction. ΔRNL will exponentially decay with the increasing separation distance. c The Hanle spin precession curve for the same device structure, showing red (black) circles for parallel (antiparallel) alignment of the FM electrodes82. The spin diffusion constant (D) and lifetime (τs) are extracted as the fitting parameters to Eq. 2. Images of b and c reprinted with permission from ref. 82 Copyright 2010 by the American Physical Society.

The graphene-based NLSV also enables a precise characterization of spin precession signals in graphene by applying an out-of-plane magnetic field and then measuring the nonlocal resistance (RNL) as a function of the magnetic field strength (the Hanle effect61,72) (Fig. 2c). Many experimental studies on the spin transport parameters of graphene have been performed by conducting the Hanle measurement in the nonlocal geometry because the Hanle spin precession curve is a strong indicator of pure spin currents in graphene. The Hanle measurement offers a more reliable way to study spin transport in graphene because the approach based on Eq. 1, requiring the NLSV device structure to be repeatedly fabricated for varying the separation distance, often encounters a significant challenge in reproducing the same FM/graphene tunnel contacts73. The spin lifetime (τs) and the spin diffusion constant (Ds) can be extracted as fitting parameters to the following equation (Eq. 2)61,72,74.

$$R_{NL} \propto \pm {\int}_0^\infty {\frac{1}{{\sqrt {4\pi D_st} }}exp\left( { - \frac{{L^2}}{{4D_st}}} \right)\cos \left( {\omega t} \right)exp\left( { - \frac{t}{{\tau _s}}} \right)dt}$$

Here, the +(-) sign is for the parallel (antiparallel) magnetization state, and ω is given by \(g\mu_BB_ \bot /\hbar\), where g is the Landé factor (g = 2.0023) and \(\hbar\) is the reduced Planck’s constant. Once τs and Ds are determined from the Hanle measurement and analysis, the corresponding spin diffusion length (LD) can be calculated by LD = (Dsτs)1/2.

Spin injection into graphene

The difficulty in utilizing the conventional semiconductor materials (e.g., silicon) for emerging spin logic applications arises from the facts that (1) their spin transport properties tend to degrade with the decreasing film thickness (i.e., poor scalability)75, and (2) the spin injection from a FM metal contact into a semiconducting channel material has been a great challenge due to a fundamental conductivity (or mode) mismatch problem76. Noticeably, a recent experimental study has reported that graphene can transport spin signals over a relatively long distance of up to 30.5 µm77 (this distance can be increased even up to 90 µm using carrier drift78). If spin injection from the FM electrode into the graphene channel can be made with a high injection efficiency, the use of graphene as a spin transport channel may be considered as a practically viable route for the development of emerging spin logic devices.

Since the first demonstration of spin transport in graphene at room temperature in 200771, spin injection efficiency into graphene has been continuously improved. Inserting a thin tunnel barrier (TB) at the FM/graphene interface has been widely used to mitigate the conductivity mismatch problem and, thus, increase the spin injection efficiency into the graphene layer. Injection efficiencies of up to ~30% have been experimentally demonstrated using a variety of TB materials such as TiO279,80, MgO81, TiO2-seeded MgO82,83, h-BN84, and Al2O385,86,87 (Table 1). More recently, even the fluorinated graphene was used as a homoepitaxial TB material to the graphene channel, achieving the tunneling spin polarization of up to 60%88. It is worth, however, noting that these TB-based approaches may still experience the above-mentioned issue of lacking the ability to reproduce the same FM/graphene tunnel contacts. More studies will need to be conducted to better quantify the spin polarization and understand what extrinsic factors (e.g., bias-dependence) impact the experimental measurements89,90,91.

Table 1 Summary of spin injection and transport properties of graphene.

Enhancing the spin injection efficiency is very important from the power perspective; the lower spin injection efficiency means that more charge currents would be pumped into the graphene channel to form a sufficiently large spin accumulation that can switch the corresponding output magnet. Most experimental studies have used Eq. 1 to estimate the spin injection efficiency parameter (η) by performing the NLSV experiment. A fundamental limitation of this approach is that the measurement accuracy of η will strongly depend on measurements of other parameters in the equation. Most notably, the spin diffusion length (LD) of graphene seems to largely vary among different studies (Table 1). Without any further optimization or new development of the graphene integration process (otherwise, the device integration process may entail exposure of graphene surfaces to organic and/or inorganic contaminants), this may impose a significant challenge in making the reliable measurement of spin transport parameters, consequently limiting the accuracy of the spin injection efficiency calculation. The next section below discusses the topic of spin transport and relaxation in the graphene spin transport channel.

Spin transport and relaxation in graphene

Table 1 summarizes the experimental measurements of graphene’s spin transport properties, with various substrate materials, FM contact electrodes, TBs, and preparation methods. Taking a close look at this table will provide several useful insights into the spin transport phenomena in graphene.

First, as seen in Fig. 3a, the use of h-BN as a substrate for graphene tends to increase the spin diffusion constant (Ds) of graphene, thus enhancing the long-distance propagation of spin signals in the graphene channel layer77,92,93. This is due to the high carrier mobility in graphene, having h-BN as the substrate; experiments have already shown that the measured Hall mobilities of graphene devices fabricated on the h-BN substrates are even comparable to the best-reported values of the suspended graphene devices94. By the Einstein relation, the high carrier mobility leads to a high diffusion constant for charge carriers. Consequently, the diffusion constant increases for spin because spin current is carried by charged particles in the NLSV. The graphene/h-BN or other 2D van der Waals (vdW) heterostructures95,96,97 may be considered as the high-performance spin transport channel.

Fig. 3: Spin transport properties of graphene.
figure 3

a Impact of graphene mobility on the spin diffusion constant (solid circles, left axis in the unit of m2/s) and lifetime (empty triangles, right axis in the unit of nanoseconds) values. It is seen that spin diffusion constant tends to increase when the graphene spin transport channel is prepared to be of high quality (high mobility). The dotted line is the fitted curve to the data points obtained from various studies in Table 1. Unlikely, the spin lifetime exhibits almost the negligible dependence on the graphene channel’s mobility. b Impact of interface engineering on the spin diffusion length (solid circles, left axis in the unit of µm) and lifetime (empty triangles, right axis in the unit of nanoseconds). Both the spin diffusion length and lifetime values are significantly enhanced upon the various interface engineering techniques used, such as annealing of the graphene surface80, the use of a large h-BN flake to prevent diffusion of solvents during the graphene transfer process77, and angle evaporation of the tunnel barrier oxides on the graphene surface82.

The phenomenon of spin relaxation in graphene can be understood by examining Table 1 on how the spin transport parameters change with other independent variables. For example, investigation of a possible relation between spin lifetime (τs) and charge carrier mobility (µ) has been a topic of great scientific interest because it would help reveal the relevant spin scattering mechanism. A linear dependence of τs on µ is suggestive of the Elliott–Yafet (E.–Y.) spin scattering mechanism98, while the inverse relation (τs ~ 1/µ) indicates the dominance of the Dyakonov–Perel (D.–P.) mechanism99. A previous study on spin relaxation in bilayer graphene films on SiO2100 observed a clear inverse dependence of τs on µ, suggesting the D.–P. model as a dominant mechanism that correlates the spin and charge transports; spin phases are randomized between collisions as electrons precess with different frequencies depending on their momenta. However, as summarized in Table 1, it is possible that both the E.–Y. and D.–P. mechanisms are simultaneously relevant80,92, or the E.–Y. scattering, where the precession frequency is conserved between collisions and the phase loss occurs only in the short time during collision, becomes a major causative factor for spin relaxation in graphene83.

Another piece of information available from Table 1 is that τs and LD values largely vary among different studies using the same TB material. This indicates that the microscopic origin for such a large discrepancy among various studies on spin transport in graphene could be best understood by examining the experimental details and setups of each study. One important conclusion from such deep analyses is that some special treatment of the graphene interface has resulted in significant improvement of the spin transport characteristics in graphene (Fig. 3b). For example, the thermal annealing technique (@ 400 °C in Ar/H2) that has been applied after transferring the CVD (chemical vapor deposition)-grown graphene layer yielded significantly higher τs and LD values of 1.2 ns and 6 µm, respectively80, as compared with other similar studies using the SiO2 substrate71,79,81,82,83,85,86,100. Furthermore, the highest values of τs and LD ever reported to date (τs of 12.6 ns and LD of 30.5 µm)77 were obtained by synergistically combining the two interfacial engineering techniques: (1) reversing the conventional NLSV fabrication order (i.e., from graphene first and then tunnel barrier, to tunnel barrier first and then graphene93), and (2) reducing the contamination of the FM/TB/graphene interfaces77. More recently, a few other works also reported spin transport in the CVD-grown graphene over the long distance of larger than 30 µm101,102. It is noted that τs and LD values, exceeding 100 ns and 100 µm, respectively, have once been reported for epitaxial graphene on SiC103. However, this experiment was conducted in the local, two-terminal geometry, limiting the measurement accuracy. Therefore, further research will need to be conducted to study if any particular interface within the FM/TB/graphene tunnel junction is responsible for significant spin dephasing or flipping. This will ultimately help suggest practical solutions to achieve the excellent spin propagation in graphene approaching the theoretically predicted values (e.g., τs ~ a few µs)104,105,106.

Inducing magnetic moments in graphene

Fundamentally, pristine graphene is strongly diamagnetic as it does not consist of d or f electrons. However, any possibility of inducing magnetic moments in graphene has been the great scientific and technological interest in the field. In particular, from the perspective of advancing the spin logic device, the existence of atomically-thin permanent magnets would mean that conventional FM thin films can be replaced by a highly scalable, novel spin injector (or detector) material. This is important because conventional FM thin films with thicknesses of less than 200 nm typically experience noticeable changes in their magnetic properties such as magnetic anisotropy107. It is also known that for the magnetic thin films even in the monolayer or few-layer range, the magnetic ordering (Curie) temperature is strongly affected by a variety of topological effects such as vacancies, substrate roughness and atomic intermixing, and incomplete crystal structures108.

Besides the recent discovery of 2D magnets49, the graphene itself has experienced significant research efforts to induce magnetic moments by extrinsic factors. Introducing light adatoms (point defects) such as hydrogen (H)109,110 or fluorine (F)111,112 has been the most popular approach to inducing localized magnetic moments in graphene. This is possible because the hydrogen or fluorine can be reversibly chemisorbed on graphene113,114, thus creating a sublattice imbalance by removing one pz orbital from the π band. According to the Lieb’s theorem115, the single hydrogen adatom can induce a quasi-localized state in graphene with magnetic moment of 1μB. A recent theoretical study even predicted that graphene in the form of quantum dots (e.g., zigzag graphene nanoflakes) can exhibit strong edge magnetism at room temperature due to the presence of localized edge states116. Despite the fact that magnetic moments in graphene have been experimentally observed by adopting the various techniques of SQUID (superconducting quantum interference device, using the Josephson junction to characterize the dependence of supercurrents on magnetic fields)111, MFM (magnetic force microscopy)117, and nonlocal110 or local112 spin transport measurements, making graphene permanently magnetic (beyond just inducing the localized magnetic moments) still remains a great challenge.

2D materials for MTJ and transistors

This section reviews recent experimental efforts to integrate 2D materials in other equally-important classes of spintronic devices—namely, MTJs and spin transistors. Owing to the unique ability of 2D materials to gain the highly homogenous atomic control over their thickness, researchers began to consider them as a critical element for these emerging spintronic devices. Many perspectives are being opened up in exploring the vast amount of 2D material family and their heterostructures, aiming to achieve specific spin device properties.


The MTJ is a modern spintronic device that can be used for processing the bit information stored in the nanomagnet29. The active area of MTJ typically consists of two FM electrodes with a thin insulating barrier sandwiched between them. The spacer layer is extremely thin (e.g., MgO in the nanometer or sub-nanometer range), leading electrons to tunnel through it. The MTJ has enabled the development of STT-MRAM (spin-transfer-torque magnetic random-access memory) for high-performance data storage and computing applications118,119,120, where spin-polarized currents exert a STT to change the magnetization direction of a nanomagnet. With its close-to-market position towards commercialization (e.g., in 2019, Everspin announced the sampling of 1 Gb STT-MRAM chips), STT-MRAM now becomes the most promising candidate to replace SRAM to bring the emerging paradigm of non-volatile embedded memory. However, further improvement of device performances and bit-cell scaling are still the great challenges due to issues related to controlling the TB and interfaces, the thermal stability, and the robustness118. In this context, 2D materials possess the potential to address some or all of these issues because they offer a precise control of the ultimate thickness while still affording sharp interfaces with other classes of materials. This will allow 2D materials to serve as an atomically-thin spacer layer in MTJ.

Graphene as a tunnel barrier

An early theoretical study on the FM/graphene interface121 triggered research on harnessing graphene as an atomically-thin spacer layer in MTJ. This was driven by the facts that (1) the in-plane lattice constant of graphene (2.46 Å) matches almost perfectly that of (111) Co or Ni (the lattice mismatch between them is only less than 1.3%), and (2) their electronic band structures in the reciprocal space overlap at the Fermi level only for one spin direction (spin filtering). Indeed, the theoretically predicted TMR ratio of the FM/graphene/FM stack was as high as about 100% owing to the perfect spin filtering effect37 assumed for the FM/graphene interface. Although Mohiuddin et al. demonstrated the first experimental work, adopting the exfoliated monolayer graphene in a Permalloy (NiFe)-based MTJ structure122, the reported TMR ratio was only 0.4%. Many other follow-up studies on the similar device stack, either using monolayer, bilayer, or few-layer graphene or other 2D materials (e.g., MoS2, WS2, h-BN), also produced the TMR ratios of less than a few percent123,124. For example, Fig. 4a shows the magnetic hysteresis with TMR of about 2% measured with the Co/graphene/NiFe structure124. This is mainly due to the integration difficulties in inserting 2D materials between oxidation-prone FM metals; unlike the conventional in situ physical vapor deposition (PVD) process that fabricates the whole MTJ stack without breaking the vacuum, most of these previous works on FM/2D material/FM made the bottom FM electrode exposed to air during the exfoliation or transfer of 2D materials.

Fig. 4: 2D Materials for MTJ.
figure 4

a The graphene-inserted magnetic tunnel junction (MTJ) device (top), and measured TMR characteristics (bottom) at temperature of 4 K and with low bias voltages (125 and 250 µV) applied124. The CVD-grown graphene is transferred onto the 20 nm NiFe FM electrode. The arrows in the TMR graph indicate the relative orientation of the NiFe (lower) and Co (higher) FM contacts. b The Co/h-BN/Fe MTJ device structure is fabricated by annealing the Fe substrate, then directly growing h-BN on the Fe catalytic substrate via the low-pressure CVD technique (top), and finally depositing the 15 nm Co top electrode by evaporation128. The conductive AFM (atomic force microscope) measurement yields the resistance values of 105 Ω for the Fe substrate and 107 Ω for the monolayer h-BN/Fe stack. The relatively high TMR ratio of about 6% is reported in this work (bottom), arising from suppression of the FM electrode oxidation due to direct CVD growth of h-BN on top of the FM electrode. c Monolayer TMD is used as the spacer layer in the ferromagnetic sandwich structure to result in the spin valve effect132,134. The NiFe/MoS2/NiFe device with monolayer MoS2 grown by CVD and transferred by the wet process (top)132 and the Co/WS2/NiFe device with the exfoliated single-layer WS2 (bottom)134 exhibit the magnetoresistance ratios of up to 0.73% and 0.47% at low temperatures, respectively. Image of a reprinted with permission from ref. 124. Copyright 2012 American Chemical Society. Image of b used from AIP Publishing with permission under a CC-BY 3.0128. Image of c (top) reprinted with permission from ref. 132. Copyright 2015 American Chemical Society. Image of c (bottom) used from Springer Nature with permission under a CC-BY 4.0134.

The existing strategic effort to prevent oxidation during the fabrication of a graphene-inserted MTJ structure falls into three major categories: inserting a thin buffer layer, the flip-transfer method, and the use of FM oxides. An example of the first category was to insert a thin noble metallic layer (e.g., Au) at the FM/graphene interface, as shown in the early work of fabricating the NiFe/graphene/Au/NiFe stack (from top to bottom) with the graphene flakes exfoliated onto the Au-coated NiFe electrode122. The flip-transfer method employed more complicated process steps of exfoliating graphene on the substrate coated by the LOR (lift-off resist)/PSS (poly (4-styrenesulfonic acid)) bi-layer, depositing the patterned NiFe electrode, separating the NiFe/graphene/LOR stack from the substrate by dissolving the PSS layer in water, and flipping it onto another SiO2/Si substrate125. However, in both the first and second approaches, the obtained TMR ratios were still limited to only about 5%. This was attributed to reduced spin polarization due to the inserted 2 nm-thick Au layer122 and the antiferromagnetic coupling between FM electrodes125, respectively. In 2014, Li et al. used a FM half-metal oxide layer of LaSrMnO3 (LSMO) to replace one of the conventional FM metal electrodes, leading to the MTJ stack of Co/graphene/LSMO126. Although LSMO is known to exhibit a very high spin polarization (about 95%), this third approach also produced negligibly small enhancement in the TMR ratio, due to contamination of the LSMO interface during the graphene transfer process.

Other 2D materials as a tunnel barrier

h-BN, a wide-bandgap 2D dielectric, can possibly better serve as a TB in MTJ due to its intrinsic insulating nature (bandgap energy, Eg ~ 6 eV). Several recent works studied the FM/h-BN/FM structures with h-BN thin films prepared by wet-transfer127 or direct CVD growth onto the FM electrode128, yielding the TMR ratios of up to 6% (Fig. 4b). The MBE (molecular beam epitaxy) technique also enabled the direct growth of h-BN onto the FM layer129, thus providing an alternative pathway to better tailoring the FM/2D material interface and ultimately creating the MTJ structure having the atomically-thin, robust TB. Despite these promising early results on the direct growth of h-BN on FM128,129, it is noted that both CVD and MBE growth of monolayer or few-layer h-BN crystals would require in-depth investigation of the parameter space for the specific FM substrate. For example, whereas graphene has already been routinely synthesized by CVD on FM catalysts such as polycrystalline Nickel (Ni), the CVD parameter space has only recently been carefully tuned (e.g., lowing the temperature down to about 450 °C) to ensure the monolayer growth of graphene on Ni130,131.

In 2015, Wang et al. introduced the CVD-grown monolayer TMD (MoS2) in MTJ132 (Fig. 4c), by using the Au-coated NiFe electrode to minimize the same oxidation issue already encountered with graphene and h-BN. In this work, the insertion of Au still led to degradation of spin polarization in MTJ, thus producing only a marginal increase in the TMR ratio (from 0.4% to 0.7%). The technique of using the FM oxides was also employed to integrate MoS2 with Fe3O4 electrodes133, but the spin signal still remained far from expectations (TMR ~0.2%). A more recent study demonstrating the MTJ device with exfoliated WS2 flakes (Fig. 4c) also resulted in a poor TMR ratio of 0.47% due to the oxidation of the bottom FM electrode134. Collectively, the use of TMD materials in MTJ seems still at early stage as compared with graphene and h-BN, requiring more advanced studies in optimizing the MTJ integration process with TMD materials.

Spin transistors

Most spintronic devices discussed so far are passive in the sense that they are not capable of amplifying an electrical signal. An active spintronic device, if successfully implemented, can find it useful in signal processing, logic circuits, etc. A transistor is the three-terminal (neglecting the body effect) active device where the electrical current flow between two terminals is modulated by a voltage (or a current) applied to the third terminal. Inspired by the seminal proposal of the Datta and Das spin modulator135, many studies have been devoted to electrical control of the spin current, which is key to implementing an emerging device concept of spin transistor136,137. In spin transistor, owing to the Rashba spin-orbit interaction (i.e., generation of an effective magnetic field by an electric field applied to the gate terminal), the spin injected from the source FM contact travels towards the drain FM contact via Larmor precession. Typically, spin transistor is normally ON (i.e., depletion mode) because no precession at zero gate voltage leads to the spin direction parallel to the drain FM. Varying the gate voltage can then alter the spin precession angle to result in a periodic modulation of the source-to-drain current. Although spin transistor promises performances of higher speed and energy-efficiency than conventional transistors138, finding the optimal device structure employing the best-suited materials for each element still requires extensive research efforts. This may open up new opportunities for 2D materials and vdW heterostructures due to their versatile crystal structures and functionalities.

TMD/graphene heterostructures as a spin channel

Yan et al. investigated the vdW heterostructure of graphene/MoS2 as a novel spin transport channel139. This study aimed to synergistically combine the unique spin-related properties of graphene and TMD; graphene allows a long-distance spin transport due to its low SOC66,67 while TMD’s strong SOC, originating from the d-orbitals of the heavy metal (the spin-orbit splitting is 148 meV in monolayer MoS2)140, holds the potential to effectively manipulate spin. In this work, graphene flakes were first exfoliated onto a highly doped silicon substrate coated with a 300 nm-thick SiO2. Then, a few-layer MoS2 flake was transferred by all-dry viscoelastic stamping141 to create a gate-tunable spin sink. Finally, several FM (Co) electrodes with TiO2 as the TB were patterned and deposited to complete the NLSV geometry. As shown in Fig. 5a, the central result was the demonstration of tuning the spin current (dictated by the nonlocal resistance measured in the NLSV) by varying the back-gate voltage (Vg) applied to the substrate. For a negative Vg of less than −10 V, the MoS2’s low sheet conductance made the spin current flow well in the adjacent graphene channel. For a positive Vg of higher than 10 V, the MoS2 sheet conductance increased to absorb much of the spin current from graphene. The authors supported this argument by estimating and comparing the spin resistances of the graphene channel and the MoS2 spin sink. Despite the lack of direct evidence of a link between the strong SOC and the proposed spin absorption mechanism in few-layer MoS2, this work featured the electrical manipulation of spin current at relatively high temperatures (up to 200 K) and the well-defined ON and OFF states. The room temperature demonstration of the electrical control of spin current in the similar MoS2/graphene heterostructure quickly followed up, with a clearer understanding of the spin transport mechanism in such a vdW heterostructure142. The great tunability of spin current was rather attributed to the gate tuning of the Schottky barrier at the MoS2/graphene interface, leading to the modulation of MoS2’s conductivity. For example, at high gate voltages applied, the reduced Schottky barrier and MoS2’s conductivity allowed spin to tunnel into MoS2, making it dephasing due to the high SOC of the MoS2 channel.

Fig. 5: 2D materials for spin transistor.
figure 5

a The device schematic of MoS2/graphene heterostructure-based spin transistor (left) and measured nonlocal resistances (Rnl) with a wide range of gate voltage biases (Vg) applied (right)139. For the nonlocal measurement, a DC current of 10 µA is injected into graphene from the Co/TiO2 tunnel contact electrode. The measured nonlocal voltage (VNL) (and thus, resistance RNL = VNL/I) varies with Vg due to the spin current (flowing in graphene) modulation by the MoS2 spin absorber (sink). b The spin tunnel FET (spin TFET) concept is proposed as an alternative way to implement spin transistor (left)146. Instead of controlling the spin current in the spin transport channel, the spin TFET harnesses electrical switching of the magnetization configuration (in much the same way as MTJ) to create the transistor output. The recently discovered 2D magnet, bilayer CrI3, is used as the active switching medium due to its spin-flip transition behavior. The fabricated device offers a large ON/OFF conductance (G) ratio approaching 400% (right). Image of a used from Springer Nature with permission under a CC-BY 4.0139. Image of b reprinted by permission from Springer Nature: ref. 146, Copyright 2019.

It is also important to note that graphene, when interfaced with TMD, exhibits a unique spin lifetime anisotropy (SLA) effect. The SLA, defined as the ratio of lifetimes for spins pointing out of the graphene plane to those pointing in the plane, was theoretically predicted to reach values of tens to hundreds in the TMD/graphene heterostructures143. This phenomenon was experimentally demonstrated in MoSe2/graphene144 and WS2 (or MoS2)/graphene145 heterostructures, showing an order of magnitude longer lifetime for out-of-plane spins (i.e., SLA~10). The SLA effect, while serving as direct evidence of strong spin-valley coupling in the TMD material (this topic will be revisited in a later section), may provide an alternative route for manipulating spin relaxation in 2D material-based spintronic devices.

2D magnets to realize a spin tunnel transistor

A very recent effort to enable spin transistor by 2D materials led to the development of a novel form of spin transistor—namely, spin tunnel field-effect transistor (spin TFET)146. In contrast to the originally proposed (Datta and Das) spin transistor concept, relying on the electrical control of spin current in the high SOC channel material, spin TFET is constructed by incorporating the emerging 2D magnetic crystal as the magnetic TB (Fig. 5b). Among many 2D magnetic materials recently discovered49,147,148,149,150,151,152, the layered magnetic insulator, CrI3, exhibits an intriguing thickness-dependent magnetic order; monolayer CrI3 is a perpendicular FM, whereas bilayer CrI3 is an antiferromagnet consisting of two FM monolayers with antiparallel magnetic moments. Importantly, under a moderate magnetic field applied (less than 1 T), the bilayer CrI3 can even be switched to FM through a spin-flip transition. Moreover, one study demonstrated that switching of bilayer CrI3 between antiferromagnetic and ferromagnetic states can be induced by an electric field153. Therefore, in this study, the vertical junction of graphene/CrI3/graphene was fabricated to examine the electrical switching behavior of CrI3, interfaced with graphene top (source) and bottom (drain) contacts (Fig. 5b). Alongside the earlier demonstrations of the spin filtering effect in CrI3, giving rise to the large TMR signal in tunnel junctions154,155,156, this work significantly advanced the spin transistor research by presenting an ambipolar transistor behavior with a high ON/OFF ratio approaching 400%. This was achieved by electrically controlling (gate tuning) the magnetization configuration of the 2D magnetic insulator, instead of directly modulating the spin current.

Spin-orbit coupling in 2D materials

Spin-orbit coupling (SOC) refers to the relativistic phenomenon of exerting a torque on the elementary particle with spin, approaching the nearby atom to “feel” the orbital motion. SOC is responsible for many inspiring physical properties of 2D materials, including the room temperature quantum Hall effect (quantization of Hall resistance under a large magnetic field) observed in monolayer graphene157,158. From the technological point of view, SOC is of great importance because it provides an effective means of generating a pure spin current even in nonmagnetic materials via the SHE (spin accumulation with no Hall voltage)159. Indeed, the room temperature SHE was recently observed in the graphene/MoS2 vdW heterostructure (Fig. 6a)160, which is best attributed to the valley-Zeeman type of SOC imprinted from TMD to graphene (i.e., the proximity-induced enhancement of SOC in graphene due to a weak hybridization with the heavy atoms in TMD)161. The charge-to-spin conversion can also occur in the graphene/TMD heterostructure via the Rashba-Edelstein effect (due to the broken out-of-plane symmetry)162, which was recently demonstrated with graphene in conjunction with the monolayer WS2163. The SOC-induced pure spin current is currently being widely researched in the field, for the purpose of developing the advanced spin-orbit torque (SOT) devices such as SOT-MRAM164 and SHO (spin Hall oscillator)165. Most early experimental studies on these devices interfaced FM with a heavy metal of large SOC (e.g., Ta, Pt) to effectively induce magnetization switching or oscillation. However, the state-of-the-art SHA (spin Hall angle, θSH, an important metric for charge current-to-spin current conversion efficiency) of heavy metals is limited to θSH of less than 0.1, with greatly varying values among different studies166. This imposes a significant challenge in developing energy-efficient, high-performance spintronic devices because the large SHA value means that a large amount of charge currents will still need to be supplied. This may offset the benefits of using spin for electronics. With the giant SOC-induced spin splitting already reported for TMD (as shown in Fig. 6b, the maximal spin splitting of the uppermost valence band at the K point is as high as 456 meV for WSe2 monolayer)140 and the relatively high SHA measured for TMD (θSH ~ 0.14 for MoS2)167 and topological insulator (TI) (θSH of up to 425 for (Bi0.5Sb0.5)2Te3 at 1.9 K)168, further research like the recent work on the FM/TI heterostructure (see Fig. 6c for the SOC-induced magnetization switching observed in Bi2Se3 films)169 can be committed to design, fabricate, and understand the ultralow-power SOT devices enabled by 2D materials. A very recent experimental work on the SOC enhancement in graphene by the proximity effect of WS2 also provided the great promise for practical device applications due to the electric-field tunability demonstrated170.

Fig. 6: Spin-orbit coupling in 2D materials.
figure 6

a The room-temperature spin Hall effect observed in graphene, in the proximity of high spin-orbit coupling 2D material (MoS2)160. A charge current applied along the graphene/MoS2 stripe (y axis) results in a spin current with out-of-plane (z axis) spin polarization in the graphene channel along the x axis. b Spin splitting (Δ in the unit of meV) of the WSe2 monolayer system, shown in the momentum space along the Γ-Κ line for the uppermost valence band (left) and the lowermost conduction band (right)140. The bottom panels show contributions of orbitals for Se (pz, px + py) and W (\(d_{3z^2 - r^2}\), \(d_{x^2 - y^2}\) + dxy). c The AFM image of a 10-QL (quintuple layers) Bi2Se3 film with a roughness of ~0.5 nm, prepared by the MBE technique (top)169. The white scale bar is 100 nm. MOKE (Magneto-optic Kerr Effect) images are used to investigate the current-driven (i.e., SOT-driven) magnetization switching in the Bi2Se3/NiFe heterostructure (bottom)169. The direction of magnetization is indicated by the white arrow. d The schematic drawing of photoinduced AHE driven by a net valley polarization of MoS2 (i.e., valley Hall effect) (top), and the source-to-drain bias (Vx) dependence of the Hall voltage (VH) measured for the monolayer MoS2 transistor device (bottom)174. R-L (L-R) indicates the quarter-wave modulation with θ = 45° (−45°) where the polarization is modulated from right-handed to left-handed (from left-handed to right-handed). No net Hall voltage is measured for the linear (half-wave s-p) modulation. No noticeable Hall voltage is observed under R-L modulation for the bilayer MoS2 case (bilayer, R-L) due to the restored inversion symmetry. Image of a reprinted with permission from ref. 160. Copyright 2019 American Chemical Society. Image of b reprinted with permission from ref. 140 Copyright 2011 by the American Physical Society. Image of c used from Springer Nature with permission under a CC-BY 4.0169. Image of d used from ref. 174. Reprinted with permission from AAAS.

It is highlighted that with semi-metallic TMD materials with reduced crystalline symmetry (e.g., WTe2171), an out-of-plane SOT can be generated to possibly induce magnetization reversal for adjacent magnetic thin films with PMA. The out-of-plane SOT that exists exclusively for currents applied along a low-symmetry axis has already been demonstrated for the (mechanically exfoliated) WTe2/ (sputtered) Permalloy bi-layer structure171. This could provide great flexibility to the SOT device design in terms of directions of charge current injection and spin current detection (unlike the conventional settings of the heavy-metal/FM or high-symmetry TMD/FM structures where the component of spin torque is required to lie within the device plane to induce most effectively the anti-damping switching). Indeed, PMA devices attract much attention recently for high-density magnetic storage device applications due to the fact that they can be scaled to small sizes while maintaining thermal stability38,39,40. This new charge-to-spin conversion mechanism through the TMD’s unique crystal symmetries may enable a new strategy for optimizing future spintronic device technologies. The possibility to achieve multi-directional charge-to-spin conversion has also been demonstrated using the low-symmetry TMD material (e.g., MoTe2172 that appears in semi-metallic, distorted 1T octahedral phases), where both the conventional (i.e., spin polarization is orthogonal to charge current) and unconventional (i.e., spin polarization is parallel to charge current) conversions were detected in the graphene-based NLSV structure172.

Using the valleys in 2D materials

Fundamentally, the phenomenon of spin polarization in conventional FM metals originates from their unique band structure. In the presence of the 3d energy band splitting with respect to the Fermi level (i.e., DOSmajority is mostly filled up being located under the Fermi level while DOSminority is mostly empty being located above the Fermi level), only charge carriers with the majority spin direction can pass through the FM layer because minority spin carriers find enough empty states in FM to scatter into. This separation of charge carriers by spin has been best studied by the anomalous Hall effect (AHE)173, where the measured Hall voltage is proportional to the FM’s magnetization (c.f., the Hall voltage is proportional to the external magnetic field in the ordinary Hall effect).

Electrons in 2D atomic crystals with a honeycomb lattice structure possess a valley degree of freedom in addition to charge and spin. Like the AHE observed in the FM material, this unique 2D system is predicted to exhibit a similar Hall effect, yet with its sign depending on the valley index. In 2014, Mak et al. reported the observation of so-called valley Hall effect (VHE) in monolayer MoS2 by illuminating the MoS2 transistor with circularly polarized light to preferentially excite electrons into a specific valley (Fig. 6d)174. In this experiment, electrons from different valleys moved in opposite directions across the MoS2 sample, with one valley being overrepresented with respect to the other. Valleytronics is a recent research paradigm that makes use of the valley of electrons; the energy bands of certain crystal lattices have two or more minima (i.e., valleys) at equal energies but at different positions in the momentum space. If the number of electrons in these valleys can be precisely controlled, thus producing so-called, valley polarization, it would be possible to create a novel form of valleytronic devices. In this context, monolayer MoS2 is a promising valleytronic material because its conduction and valence-band edges have two energy-degenerate valleys at the corners (K points) of the first Brillouin zone175.

Valley polarization in MoS2

In much the same way that spintronic devices require spin polarization, valley polarization (non-equilibrium charge carrier imbalance between valleys) serves as the essential mechanism to realize the future valleytronic devices. Valley polarization has previously been demonstrated using strain176,177 or magnetic fields178,179,180. However, none of these approaches allowed dynamic control of charge carriers. More importantly, most of these studies relied on a special device geometry such as quantum well (e.g., AlAs-based two-dimensional electron system). In 2012, Cao et al. performed first-principles calculations on monolayer MoS2, suggesting that valley polarization is achievable via valley-selective circular dichroism (the two valleys absorb left- and right-handed photons differently) arising from its unique symmetry181. In monolayer MoS2, each Mo sits in the center of a trigonal prismatic cage formed by six sulfur atoms, thus forming the stable honeycomb lattice structure with the broken inversion symmetry. This is the essential condition for valley-selective circular dichroism (VSCD), which is further enhanced by the strong SOC associated with the transition metal’s d orbitals. Experimental demonstrations of valley polarization in monolayer MoS2 quickly followed up using the optical pumping technique182,183.

More recently, valley polarization was generated in MoS2 by the electrical method, either in a gated bilayer MoS2 transistor structure (the perpendicular gate bias broke the inversion symmetry in bilayer MoS2)184 or using a nonlocal Hall bar device geometry with monolayer MoS2185. It is important to note that the inverse valley Hall effect also made it possible to detect valley polarization in MoS2 electrically185. These studies provide researchers with rich insights on generating, controlling, and detecting valley polarization in MoS2, which can ultimately form the basis of developing the future valleytronic devices based on 2D materials.

Spin-valley coupling in TMD

The broken inversion symmetry together with strong SOC directly leads to spin-valley coupled properties (spin and valley indices of charge carriers are coupled) in monolayer MoS2 and other group-VI dichalcogenides175. From the device perspective, this means that monolayer TMD can provide a unique integrated platform to use valley for better manipulating spin. The strong spin-valley coupling in these 2D materials has two specific consequences for device applications. First, spin and valley relaxation are suppressed at the valence-band edges as flipping each index alone is prohibited by the valley-contrasting spin splitting (0.1–0.5 eV) caused by inversion symmetry breaking. For spintronic devices requiring a long-range spin transport, this large spin-valley coupling may offer a solid mechanism to enhance the spin propagation by controlling the polarization of valleys in 2D materials. Secondly, the valley-dependent optical selection rule (e.g., VSCD181) also becomes spin-dependent, thus possibly resulting in photon-induced spin and valley Hall effects. This may be used to produce long-lived spin and valley accumulations on monolayer TMD’s boundaries, which is key to the operation of spintronic and valleytronic devices.


The last few decades of nanomagnetism and spintronics research featured remarkable advances in the development of information storage devices based on multilayers of magnetic metals and insulators. However, high-volume information processing and communication devices, either enabled or assisted by spin, still have a long way to go before commercialization, mainly due to relatively slow progress in semiconductor spintronics (in particular, searching for right materials in right places). Recently, researchers have begun to realize the great potential of 2D materials for advanced spintronic device applications. In this section, several challenges and future perspectives are discussed in terms of material development and device integration for 2D spintronics.

2D magnets

Ferromagnetic order in 2D materials is a coveted property that would allow fundamental studies of spin-dependent properties in low-dimensional nanostructures and further enable novel spintronic device applications. Cr-based vdW crystals such as CrI3 and Cr2Ge2Te6 were reported to be intrinsically ferromagnetic49,147, showing the great promise to build practical spintronic devices through the use of atomically-thin ferromagnetic materials. The discovery of intrinsic ferromagnetism in these layered structures attracted significant attention in the field because making 2D materials magnetic by defect engineering109,110,111,112 may not be practically applicable (i.e., the induced magnetism is quite weak and lacks robustness). However, the low Curie temperature (Tc) of these materials (Tc is about 45 K for monolayer CrI3. For a complete list of Tc values reported for a variety of 2D magnetic materials, readers find a recent review article186 a useful resource) still remained a great challenge in adopting them for practical room-temperature operation of spintronic devices. More recently, room-temperature ferromagnetism was observed in the VSe2/MoS2 vdW heterostructure149, MnSex epitaxial films150, Fe3GeTe2151, and transition metal dihydride (MH2) monolayers152, leading the possible pathway to offer spin-polarized carriers through 2D materials at room temperature. For the purpose of electrical spin injection, 2D magnets with half-metallicity (i.e., acts as a metal in one spin direction and an insulator in the other direction)152 may be worth more research efforts because it offers nearly perfect spin polarization, making it possible to achieve a very high spin injection efficiency in spin transistors or logic switches. For metallic spintronic devices such as MTJs, all-2D MTJ structures (the FM electrodes and the TB are all made of 2D materials) may be constructed by combining with the recent progress in integrating 2D materials as the atomically-thin TB123. The device functionality based on 2D magnets already started to be reported, for example, by demonstrating the SOT-induced magnetization switching in Fe3GeTe2187.

Interfacing 2D materials with FM

Although, in principle, 2D materials are expected to be readily interfaced with other classes of materials, including FMs, largely due to its excellent flexibility and other mechanical properties188, one practical challenge is to prevent oxidation or any other surface contamination of the substrate material during the integration process. Recalling that many existing strategies developed for addressing this issue suffered from degradation of spin polarization in the MTJ structure122,125,126,127,132,133,134, other new approaches may be implemented to better integrate 2D materials with FM thin films. Indeed, researchers have already started growing graphene or h-BN directly on top of the FM catalytic substrate using the carefully-controlled CVD or MBE technique128,129,130,131 instead of relying on ex-situ processes such as exfoliation or transfer. Direct growth of TMD materials onto FM substrates has not been demonstrated yet. A recently developed intercalation technique of inserting the FM layer (Co) into an already formed graphene/substrate interface may provide another path for creating an excellent FM/graphene interface189. However, the impact of these new processes on the specific spintronic device performance still remains unknown, requiring more advanced studies of applying novel integration processes to create and evaluate the 2D spintronic devices.

Implementation of spin logic

Using spin as a state variable in logic devices provides several advantages over conventional charge-based information processing technologies; non-volatility, faster and more energy-efficient data processing, and increased integration densities. Especially, theorists predict that in spin-based computing, if the rate of spin relaxation is maintained much slower than the switching rate, energy dissipation can be even less than the thermal limit (« 10−21 J/bit)190. If properly implemented, spintronic computing could better satisfy the rapidly increasing performance needs of emerging abundant-data applications191. All-spin logic (ASL)51 has been considered as one of the most effective ways to implement spin logic with the highest energy-efficiency because both its internal and terminal quantities are spin, requiring no spin-to-charge conversion. To date, despite its great promise, the experimental demonstration of ASL has not been made yet. As already discussed in this article, 2D materials have the potential to address some of the challenges of ASL, such as developing the long-range, high injection-efficiency spin transport channel material.

The concept of ASL, however, continues to evolve towards other device approaches that are more easily implementable or can better represent the emerging computing paradigm. For example, the charge-coupled spin logic (CSL)58,192 uses robust charge currents for long-distance transport as terminal quantities. In the original proposal of CSL, the SHE was suggested as the write mechanism while MTJ served as the read unit. Recently, the CSL device was experimentally implemented based on a vertically-coupled pair of in-plane and PMA nanomagnets, clearly demonstrating its functionality as a logic gate and a true random number generator58. More recently, the newer concept of a probabilistic bit (p-bit)193 was introduced as an intermediate computing scheme between the conventional bits of digital electronics and the emerging quantum bit (qubit). In the probabilistic spin logic (PSL), low-barrier nanomagnets (i.e., engineered to be purposely unstable) provide a natural physical representation for p-bit, which is the classical entity that fluctuates rapidly between 0 and 1. The PSL may open up new application spaces in an emerging computing hardware domain because of their similarity. For example, the p-bit MTJ, involving the low-barrier nanomagnet as a free layer, can be easily combined with the MOS transistor to create the conventional 1T1R (one transistor-one resistor) circuitry. This could offer an efficient hardware platform to accelerate deep learning computation like the binary stochastic neuron194. A new spin logic device concept was also proposed by combining the spin-orbit coupling (for reading) and the magnetoelectric effect (for writing), showing that this magnetoelectric spin-orbit logic exhibits superior energy efficiency than the CMOS technology (by a factor of 10 to 30)195. This may benefit greatly from the unique spin properties (large spin Hall angles, long spin lifetime) of 2D materials and their heterostructures196. Looking further into how 2D materials can better enable or advance these newer forms of the spin logic device will potentially pave the way for the post-silicon spintronics era.