IOP PUBLISHING JOURNAL OF PHYSICS D: APPLIED PHYSICS

J. Phys. D: Appl. Phys. 40 (2007) R337–R354 doi:10.1088/0022-3727/40/21/R01

TOPICAL REVIEW

Giant tunnel magnetoresistance inmagnetic tunnel junctions with acrystalline MgO(0 0 1) barrierS Yuasa1 and D D Djayaprawira2

1 National Institute of Advanced Industrial Science and Technology (AIST),Nanoelectronics Research Institute, 1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, Japan2 Electron Device Equipment Division, Canon ANELVA Corporation, 5-8-1 Yotsuya,Fuchu-shi, Tokyo 183-8508, Japan

E-mail: yuasa-s@aist.go.jp

Received 22 May 2006, in final form 15 August 2007Published 19 October 2007Online at stacks.iop.org/JPhysD/40/R337

AbstractA magnetic tunnel junction (MTJ), which consists of a thin insulating layer(a tunnel barrier) sandwiched between two ferromagnetic electrode layers,exhibits tunnel magnetoresistance (TMR) due to spin-dependent electrontunnelling. Since the 1995 discovery of room-temperature TMR, MTJs withan amorphous aluminium oxide (Al–O) tunnel barrier have been studiedextensively. Al–O-based MTJs exhibit magnetoresistance (MR) ratios up toabout 70% at room temperature (RT) and are currently used inmagnetoresistive random access memory (MRAM) and the read heads ofhard disk drives. MTJs with MR ratios significantly higher than 70% at RT,however, are needed for next-generation spintronic devices. In 2001first-principle theories predicted that the MR ratios of epitaxial Fe/MgO/FeMTJs with a crystalline MgO(0 0 1) barrier would be over 1000% because ofthe coherent tunnelling of fully spin-polarized �1 electrons. In 2004 MRratios of about 200% were obtained in MTJs with a single-crystalMgO(0 0 1) barrier or a textured MgO(0 0 1) barrier. CoFeB/MgO/CoFeBMTJs for practical applications were also developed and found to have MRratios up to 500% at RT. MgO-based MTJs are of great importance not onlyfor device applications but also for clarifying the physics of spin-dependenttunnelling. In this article we introduce recent studies on physics andapplications of the giant TMR in MgO-based MTJs.

1. Introduction

1.1. Tunnel magnetoresistance (TMR) effect

The resistance of a magnetic tunnel junction (MTJ), whichconsists of a thin insulating layer (a tunnel barrier) sandwichedbetween two ferromagnetic (FM) metal layers (electrodes),depends on the relative magnetic alignment (parallel orantiparallel) of the electrodes as shown in figure 1. The

tunnelling resistance R of the junction is lower when themagnetizations are parallel (figure 1(a)) than it is when themagnetizations are antiparallel (figure 1(b)). That is, RP <

RAP. This change in resistance with the relative orientationof the two magnetic layers, called the TMR effect, is one ofthe most important phenomena in spintronics. The size ofthis effect is measured by the fractional change in resistance(RAP–RP)/RP, which is called the magnetoresistance ratio (MRratio).

0022-3727/07/210337+18$30.00 © 2007 IOP Publishing Ltd Printed in the UK R337

Topical Review

Figure 1. Schematic illustration of the TMR effect in a MTJ. (a)Magnetizations in the two electrodes are aligned parallel (P state).(b) Magnetizations are aligned antiparallel (AP state). D1↑ and D1↓,respectively, denote the density of states at EF for the majority-spinand minority-spin bands in electrode 1, and D2↑ and D2↓respectively denote the density of states at EF for the majority-spinand minority-spin bands in electrode 2.

The TMR effect was first observed in 1975 by Julliere [1],who found that a Fe/Ge–O/Co MTJ exhibited a MR ratio of14% at 4.2 K. Although for more than a decade it receivedlittle attention because it was not observed at room temperature(RT), it attracted renewed attention after the discovery of giantmagnetoresistance (GMR) in metallic magnetic multilayers inthe late 1980s [2,3]. Because the GMR effect was recognizedto be applicable to magnetic sensor devices such as the readheads of hard disk drives (HDDs), extensive experimental andtheoretical efforts were devoted to increasing the MR ratioat RT. In 1995 Miyazaki et al [4] and Moodera et al [5] madeMTJs with amorphous aluminium oxide (Al–O) tunnel barriersand 3d ferromagnetic electrodes and obtained MR ratios ashigh as 18% at RT. Because these room-temperature MR ratioswere the highest yet reported for a ferromagnetic/nonmagnetic(NM)/ferromagnetic trilayer structure called a (pseudo-)spin-valve structure, the TMR effect attracted a great deal ofattention. Although room-temperature MR ratios, shown infigure 2, have been increased to about 70% by optimizingthe ferromagnetic electrode materials and the conditions forfabricating the Al–O barrier [6], they are still lower thanneeded for many applications of spintronic devices. High-density magnetoresistive random-access-memory (MRAM)cells (figure 3), for example, will need to have MR ratios thatare higher than 150% at RT, and the read head in the next-generation ultrahigh-density HDD will need to have both ahigh MR ratio and an ultralow tunnelling resistance. The MRratios of the conventional Al–O-based MTJs are simply nothigh enough for next-generation device applications.

In 2001 first-principle calculations predicted that epitaxialMTJs with a crystalline magnesium oxide (MgO) tunnel barrierwould have MR ratios of over 1000%, and in 2004 MR ratiosof about 200% were obtained at RT in MTJs with a crystalline

Figure 2. History of improvement in MR ratio at room temperature(RT) for MTJs with an amorphous Al–O barrier.

Figure 3. A (a) circuit diagram and (b) typical cross-sectionalstructure of a MRAM cell. (c) A typical cross-sectional structure ofa MTJ for practical applications. (d) A typical magnetoresistancecurve of a MTJ and the definition of MR ratio.

MgO(0 0 1) barrier. The huge TMR effect in MgO-based MTJsis now called the giant TMR effect and is of great importance notonly for device applications but also for clarifying the physicsof spin-dependent tunnelling. In this review, we present anoverview of the giant TMR effect in MgO-based MTJs. Inthe following two subsections of this Introduction, we providefurther background information about the TMR effect in MTJswith an amorphous Al–O barrier. In section 2 we explainthe theoretical basis of the TMR effect in MgO-based MTJs,and in sections 3 and 4 we discuss experimental studies onfully epitaxial MgO-based MTJs. In section 5 we explain thestructure and fabrication of CoFeB/MgO/CoFeB MTJs, andin section 6 we discuss device applications of the giant TMReffect in MgO-based MTJs.

R338

Topical Review

Figure 4. (a) Estimation of spin polarization (P ) from the observedMR ratio by using Julliere’s model. (b) Direct measurement of P byusing ferromagnet/Al–O/superconductor tunnel junction. (c)Theoretical definition of P .

1.2. Julliere’s model and spin polarization

Julliere proposed a simple phenomenological model, in whichthe TMR effect is due to spin-dependent electron tunnelling[1]. According to this model the MR ratio of an MTJ canbe expressed in terms of the spin polarizations P of theferromagnetic electrodes

MR = 2P1P2/(1 − P1P2), (1)

where

Pα ≡ [Dα↑(EF) − Dα↓(EF)]/[Dα↑(EF) + Dα↓(EF)];α = 1, 2. (2)

Here Pα is the spin polarization of a ferromagnetic electrode,and Dα↑(EF) and Dα↓(EF) are, respectively, the densities ofstates (DOS) of the electrode at the Fermi energy (EF) forthe majority-spin and minority-spin bands (see figure 1). InJulliere’s model spin polarization is an intrinsic property of anelectrode material. When an electrode material is NM, P = 0.When the DOS of the electrode material is fully spin-polarizedat EF, |P | = 1.

The spin polarization of a ferromagnet at lowtemperature can be directly measured using ferromagnet/Al–O/superconductor tunnel junctions [7]. Measured this way,the spin polarizations of 3d ferromagnetic metals and alloysbased on iron (Fe), nickel (Ni) and cobalt (Co) are alwayspositive and usually between 0 and 0.6 at low temperaturesbelow 4.2 K [7, 8]. The MR ratios estimated from Julliere’smodel (equation (1)) using these measured P values agreerelatively well with the MR ratios observed experimentally inMTJs (figure 4), but the theoretical values of P (equation (2))obtained from band calculations do not agree with themeasured spin polarizations and the MR ratios observedexperimentally (figure 4). Even the signs of P often differbetween theoretical values and experimental results. Forexample, according to the band structures of Co and Ni, these

metals have negative spin polarizations, but the P valuesobserved experimentally for these metals are positive. Thisdiscrepancy, one of the most fundamental questions withregard to the TMR effect, is discussed in the followingsubsection.

Julliere’s model with the spin polarizations measuredexperimentally (0 < P < 0.6) yields a maximum MRratio of about 100% at low temperatures. A MR ratio ofabout 70% at RT is therefore close to the Julliere limit forthe 3d-ferromagnetic-alloy electrodes if a reduction in P dueto thermal spin fluctuations at finite temperatures is takeninto account. One way to obtain a MR ratio significantlyhigher than 70% at RT is to use as electrodes special kindsof ferromagnetic materials called half metals, which have afull spin polarization (|P | = 1) and are therefore theoreticallyexpected to give MTJs huge MR ratios (up to infinity, accordingto Julliere’s model). Some candidate half metals are CrO2,Heusler alloys such as Co2MnSi, Fe3O4 and manganeseperovskite oxides such as La1−xSrxMnO3. Very high MRratios, above several hundred percent, have been obtainedat low temperature in La1−xSrxMnO3/SrTiO3/La1−xSrxMnO3

MTJs [9] and Co2MnSi/Al–O/Co2MnSi MTJs [10]. However,such high MR ratios have never been observed at RT for half-metal electrodes3. Another way to obtain a very high MR ratiois to use coherent spin-dependent tunnelling in an epitaxialMTJ with a crystalline tunnel barrier such as MgO(0 0 1). Thatis the main subject of this article.

1.3. Incoherent tunnelling through an amorphous Al–Obarrier

Before going into the details of coherent tunnelling througha crystalline MgO(0 0 1) barrier, we explain an incoherenttunnelling process through the amorphous Al–O tunnel barrier.Tunnelling in a MTJ with an amorphous Al–O barrier isillustrated schematically in figure 5(a), where the top electrodelayer is Fe(0 0 1) as an example of a 3d ferromagnet. VariousBloch states with different symmetries of wave functions existin the electrode. Because the tunnel barrier is amorphous,there is no crystallographic symmetry in the tunnel barrier.Because of this nonsymmetrical structure, Bloch states withvarious symmetries can couple with evanescent states in Al–O and therefore have finite tunnelling probabilities. Thistunnelling process can be regarded as an incoherent tunnelling.In 3d ferromagnetic metals and alloys, Bloch states with�1 symmetry (spd hybridized states) usually have a largepositive spin polarization at EF, whereas Bloch states with�2 symmetry (d states) often have a negative spin polarization

3 MR ratios of above 200% have recently been observed at room temperaturein fully epitaxial MTJs with a MgO(0 0 1) tunnel barrier and Heusler-alloyelectrodes (e.g. Tezuka N et al 2006 Appl. Phys. Lett. 89 252508). This largeTMR effect, however, is thought to originate from the coherent tunnelling ina crystalline MgO(0 0 1) barrier rather than from the half-metallic nature ofthe electrodes. Note that, as described in sections 3 and 5, when combinedwith a crystalline MgO(0 0 1) barrier, even simple ferromagnetic electrodessuch as bcc Fe, Co and CoFeB yield MTJs with MR ratios from 180% to500% at RT [25, 28, 38, 39]. By the way, Sakuraba et al [10] observed a MRratio of 570% at low temperature in MTJs with an amorphous Al–O barrierand Heusler-alloy electrodes. They also observed a feature characteristic ofa spin-dependent band gap in the tunnelling spectra for those MTJs. Thisgiant TMR effect at low temperature is therefore thought to be due to thehalf-metallic nature of Heusler-alloy electrodes.

R339

Topical Review

Figure 5. Schematic illustrations of electron tunnelling through (a)an amorphous Al–O barrier and (b) a crystalline MgO(0 0 1) barrier.

at EF. Julliere’s model assumes that tunnelling probabilitiesare equal for all the Bloch states in the electrodes. Thisassumption corresponds to a completely incoherent tunnelling,in which none of the momentum and coherency of Bloch statesis conserved. This assumption, however, is not valid even inthe case of an amorphous Al–O barrier. Although the spinpolarization P defined by equation (2) is negative for Co andNi, the P observed experimentally for these materials whenthey are combined with an Al–O barrier is positive [7,8]. Thisdiscrepancy indicates that the tunnelling probability in actualMTJs depends on the symmetry of each Bloch state. Theactual tunnelling process can be explained in the followingway. The �1 Bloch states with larger P are considered tohave higher tunnelling probabilities than the other Bloch states[11, 12]. This results in a positive net spin polarization of theferromagnetic electrode. Because the other Bloch states, suchas �2 states (P < 0), also contribute to the tunnelling current,the net spin polarization of the electrode is reduced below 0.6in the case of the usual 3d ferromagnetic metals and alloys.If only the highly spin-polarized �1 states coherently tunnelthrough a barrier (figure 5(b)), a very high spin polarization oftunnelling current and thus a very high MR ratio are expected.Such an ideal coherent tunnelling is theoretically expected inan epitaxial MTJ with a crystalline MgO(0 0 1) tunnel barrier(see section 2). It should be noted here that the actual tunnellingprocess through the amorphous Al–O barrier is considered tobe an intermediate process between the completely incoherenttunnelling represented by Julliere’s model and the coherenttunnelling illustrated in figure 5(b).

2. Theory of coherent tunnelling through acrystalline MgO(0 0 1) barrier

As shown in figure 6, a crystalline MgO(0 0 1) barrierlayer can be epitaxially grown on a bcc Fe(0 0 1) layerwith a relatively small lattice mismatch of about 3%.This amount of lattice mismatch can be absorbed bylattice distortions in the Fe and MgO layers and/or bydislocations formed at the interface. Coherent tunnellingtransport in epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ isillustrated schematically in figure 5(b). In the case of idealcoherent tunnelling, Fe �1 states are theoretically expectedto dominantly tunnel through the MgO(0 0 1) barrier by thefollowing mechanism [13,14]. For the k‖ = 0 direction ([0 0 1]direction in this case), in which the tunnelling probability is the

highest, there are three kinds of evanescent states (tunnellingstates) in the band gap of MgO(0 0 1): �1, �5 and �2’. Itshould be noted here that although conventional theories oftenassume tunnelling states to be plane waves, they actuallyhave specific orbital symmetries. When the symmetries oftunnelling wave functions are conserved, Fe �1 Bloch statescouple with MgO �1 evanescent states, as shown in figure 7(a).Figure 7(b) shows the partial DOS (obtained by first-principlecalculations) for the decaying evanescent states in a MgObarrier layer in the case of parallel magnetic alignment [13].Of these states, the �1 evanescent states have the slowestdecay (i.e. the longest decay length). The dominant tunnellingchannel for the parallel magnetic state is Fe �1 ↔ MgO�1 ↔ Fe �1. Band dispersion of bcc Fe for the [0 0 1](k‖ = 0) direction is shown in figure 8(a). The net spinpolarization of Fe is small because both majority-spin andminority-spin bands have many states at EF, but the Fe �1

band is fully spin-polarized at EF (P = 1). A very large TMReffect in the epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ istherefore expected when �1 electrons dominantly tunnel. Itshould also be noted that a finite tunnelling current flows evenfor antiparallel magnetic states. Tunnelling probability as afunction of k‖ wave vectors (kx and ky) is shown in figure 9 [13].For the majority-spin conductance in the parallel magneticstate (P state) (figure 9(a)), tunnelling takes place dominantlyat k‖ = 0 because of the coherent tunnelling of majority-spin�1 states. For the minority-spin conductance in the P state(figure 9(b)) and the conductance in antiparallel magnetic state(AP state) (figure 9(c)), spikes of tunnelling probability appearat finite k‖ points called hot spots. This hot-spot tunnelling isresonant tunnelling between interface resonant states [13,14].Although a finite tunnelling current flows in the AP state, thetunnelling conductance in the P state is much larger than thatin the AP state, making the MR ratio very high.

It should be noted that the �1 Bloch states are highlyspin-polarized not only in bcc Fe(0 0 1) but also in manyother bcc ferromagnetic metals and alloys based on Fe andCo (e.g. bcc Fe–Co, bcc CoFeB and some of the Heusleralloys). As an example, band dispersion of bcc Co(0 0 1) (ametastable structure) is shown in figure 8(b). The �1 statesin bcc Co, like those in bcc Fe, are fully spin-polarized atEF. According to first-principle calculations, the TMR of aCo(0 0 1)/MgO(0 0 1)/Co(0 0 1) MTJ is even larger than thatof an Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ [16]. A very largeTMR should be characteristic of MTJs with 3d-ferromagnetic-alloy electrodes with bcc(0 0 1) structure based on Fe and Co.Note also that very large TMR is theoretically expected notonly for the MgO(0 0 1) barrier but also for other crystallinetunnel barriers such as ZnSe(0 0 1) [17] and SrTiO3(0 0 1)[18]. Large TMR has, however, never been observed in MTJswith crystalline ZnSe(0 0 1) or SrTiO3(0 0 1) barriers becauseof experimental difficulties in fabricating high-quality MTJswithout pin-holes and interdiffusion at the interfaces.

3. Giant TMR effect in epitaxial MTJs with acrystalline MgO(0 0 1) barrier

Since the theoretical predictions of very large TMReffects in Fe/MgO/Fe MTJs [13, 14], there have beenseveral experimental attempts to fabricate fully epitaxial

R340

Topical Review

Figure 6. Crystallographic relationship and interface structure of epitaxial bcc Fe(0 0 1)/NaCl-type MgO(0 0 1): (a) top view and (b)cross-sectional view. aFe and aMgO denote the lattice constants of bcc Fe and NaCl-type MgO unit cells.

Figure 7. (a) Coupling of wave functions between the Bloch states in Fe and the evanescent states in MgO for the k‖ = 0 direction. (b)Tunnelling DOS of majority-spin states for k‖ = 0 in Fe(0 0 1)/MgO(0 0 1)(8 ML)/Fe(0 0 1) with parallel magnetic state. (Adaptedfrom [13].)

Figure 8. (a) Band dispersion of bcc Fe in the [0 0 1] (�–H)direction. (b) Band dispersion of bcc Co in the [0 0 1] (�-H)direction. (Redrawn from Bagayako et al [15].) Thin black and greylines respectively represent majority-spin and minority-spin bands.Thick black and grey lines respectively represent majority-spin andminority-spin �1 bands. EF denotes Fermi energy. For easycomparison of the relative levels of EF in bcc Fe and bcc Co withrespect to the majority-spin �1 band, the bottom edges of themajority-spin �1 bands in (a) and (b) are aligned at the same energylevel.

Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJs [19–21]. Bowen et alwere the first to obtain a relatively high MR ratio inFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJs at RT (30%) [20], but theroom-temperature MR ratios obtained in MgO-based MTJsdid not exceed the highest one obtained in Al–O-based MTJs(70%). The main difficulty at the early stage of experimentalattempts was the fabrication of an ideal interface structure likethat shown in figure 6(b). It was experimentally observedthat Fe atoms at the Fe(0 0 1)/MgO(0 0 1) interface were easilyoxidized [22]. Results of first-principle calculations for theideal interface and the oxidized interface are shown in figure 10[23]. At the ideal interface (figure 10(a)), where there are noO atoms in the first Fe monolayer at the interface, the Fe �1

Bloch states effectively couple with the MgO �1 evanescentstates in the k‖ = 0 direction, which results in the coherenttunnelling of �1 states and thus the very large TMR effect. Atthe oxidized interface (figure 10(b)), where there are excessoxygen atoms in the interfacial Fe monolayer, the Fe �1 statesdo not couple with the MgO�1 states effectively. This preventscoherent tunnelling of �1 states and significantly reducesthe MR ratio. Coherent tunnelling is thus very sensitiveto the structure of barrier/electrode interfaces. Oxidation ofeven a monolayer at the interface significantly suppresses theTMR effect.

R341

Topical Review

Figure 9. Tunnelling probability in a Fe(0 0 1)/MgO(0 0 1)(4 ML)/Fe(0 0 1) MTJ as a function of kx and ky wave vectors. (Adaptedfrom [13].) (a) Majority-spin conductance in the parallel magneticstate (P state), (b) minority-spin conductance in the P state, (c)conductance in the antiparallel magnetic state (AP state).

Figure 10. Partial density of states at EF due to the majority-spin�1 states near the (a) ideal and (b) oxidized Fe(0 0 1)/MgO(0 0 1)interface. (Redrawn from Zhang et al [23].)

We fabricated high-quality fully epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJs by using MBE growth under ultrahighvacuum [24, 25]. Cross-sectional transmission electronmicroscope (TEM) images of the MTJ are shown in figure 11,where single-crystal lattices of MgO(0 0 1) and Fe(0 0 1) areclearly evident. Magnetoresistance curves of the epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ are shown in figure 12(a).We obtained MR ratios up to 180% at RT (① and ②‚ infigure 13), and these were the first RT MR ratios higher thanthe highest RT MR ratio obtained with an Al–O-based MTJ.The key to obtaining such high MR ratios is thought to be

Figure 11. Cross-sectional transmission electron microscope (TEM)images of an epitaxial Fe(0 0 1)/MgO(0 0 1)(1.8 nm)/Fe(0 0 1) MTJ.(Reprinted from [25].) (b) is a magnified view of part of (a). Thevertical and horizontal directions respectively correspond to theMgO[0 0 1] (Fe[0 0 1]) axis and MgO[1 0 0] (Fe[1 1 0]) axis.

the production of clean Fe/MgO interfaces without oxidizedFe atoms. X-ray absorption spectroscopy (XAS) and x-raymagnetic circular dichroism (XMCD) studies revealed thatnone of the Fe atoms adjacent to the MgO(0 0 1) layer areoxidized, indicating that there are no oxygen atoms in theinterfacial Fe monolayer [26]. Parkin et al fabricated MTJswith a highly oriented polycrystalline (or textured) MgO(0 0 1)barrier by using sputtering deposition on a SiO2 substrate witha TaN seed layer that was used to orient the entire MTJ stack inthe (0 0 1) plane [27]. They obtained MR ratios up to 220% atRT (③ in figure 13). It should be noted that the fully epitaxialMTJs and the textured MTJs are basically the same from amicroscopic point of view. We also fabricated fully epitaxialCo(0 0 1)/MgO(0 0 1)/Co(0 0 1) MTJs with metastable bccCo(0 0 1) electrodes by using MBE growth and obtained RTMR ratios of above 400% (figure 12(b)) [28]. The very largeTMR effect in MgO-based MTJs is called the giant TMR effect.

4. Other phenomena observed in epitaxial MTJswith a crystalline MgO(0 0 1) barrier

It is difficult to investigate the detailed mechanisms of TMRby using the amorphous Al–O barrier because of its structuraluncertainty, and there has been no significant progress inunderstanding the physics of the TMR effect since Julliereproposed the phenomenological model [1]. Unlike Al–O-based MTJs, epitaxial MTJs with a single-crystal MgO(0 0 1)barrier are a model system for studying the physics of spin-dependent tunnelling because of their well-defined crystallinestructure with atomically flat interfaces. As explained below,interesting phenomena other than the giant TMR effect havebeen observed in epitaxial MgO-based MTJs. Clarifying theirmechanisms will lead to a deeper understanding of the physicsof spin-dependent tunnelling.

4.1. Oscillatory TMR effect with respect to MgO barrierthickness

The MR ratio of epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJshas been found to oscillate with respect to MgO barrier

R342

Topical Review

Figure 12. Magnetoresistance curves (at a bias voltage of 10 mV) measured at RT and 20 K for (a) an epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ (reprinted from [25]) and (b) an epitaxial Co(0 0 1)/MgO(0 0 1)/Co(0 0 1) MTJ with metastable bccCo(0 0 1) electrodes (reprinted from [28]). Arrows represent magnetization alignments. In these MTJs the top ferromagnetic electrode layeris exchange-biased by an antiferromagnetic Ir–Mn layer.

Figure 13. History of improvement in MR ratio at roomtemperature (RT).

thickness tMgO as described below [25, 29]. We made a high-quality fully epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ filmwith a wedge-shaped MgO barrier layer by using MBE growth.A cross-section of the film is shown in figure 14(a). Thefilm with a wedge-shaped MgO layer was made into MTJswith a junction area A of 36 µm2 (figures 14(b)–(d)) by usinghigh-precision microfabrication processes [29] that suppressedsample-to-sample variation in junction area A and thus resultedin small sample-to-sample variation in the transport properties.Figure 15 shows the tMgO-dependence of a resistance–area(RA) product: tunnelling resistance for a unit junction area.The tunnelling resistance increases exponentially with respectto the barrier thickness (tMgO), which is a typical characteristicof tunnelling. The tMgO-dependence of the MR ratio is shownin figure 16(a). Surprisingly, the MR ratio is seen to oscillateas a function of tMgO. As shown in the solid blue line infigure 16(a), this oscillation was fitted relatively well by asingle-period oscillation function (a simple cosine curve) withan oscillation period (λ) of 3.17 Å plus a background curve.The small discrepancy between the observed MR ratios andthe fitting curve is discussed later. It should be noted that suchan oscillatory barrier thickness dependence of MR has neverbeen observed for MTJs with an amorphous Al–O barrier. It

Figure 14. (a) Cross-section of a fully epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ film with a MgO layer that hasa wedge-shaped cross-section. (b)–(d) Top-view photographs ofFe/MgO/Fe MTJs prepared by high-precision microfabricationtechniques: (b) overall view of all MTJs fabricated on the samesubstrate, (c) close-up of one MTJ with contact pads and (d) opticalmicroscopic image of an MTJ with a junction area of 36 µm2.(Reprinted from [29].)

might be thought that the oscillation is a result of the layer-by-layer epitaxial growth of MgO(0 0 1), in which the growth ofone monolayer is almost completed before the growth of a newmonolayer begins. This growth could cause an oscillation inthe MR ratio because the interface morphology (atomic stepdensity) changes periodically, layer by layer. This cannot bethe origin of the observed oscillation, however, because theoscillation period (3.17 Å) is not the thickness of a monoatomicMgO(0 0 1) layer (2.1 Å).

Butler et al proposed a model of interference betweentunnelling states [13]. Interference between the two evanescentstates at EF in MgO that correspond to �1 and �5 at k‖ = 0could cause an oscillation of tunnelling transmittance as afunction of tMgO as follows. These states have complex wavevectors, the perpendicular components (z-components) ofwhich are expressed as k1 = kr

1+iκ1 and k2 = kr2+iκ2. Nonzero

R343

Topical Review

Figure 15. MgO-thickness (tMgO)-dependence of resistance-area(RA) product in parallel magnetic state (P state) (open circles) andantiparallel magnetic state (AP state) (solid circles) at 20 K forepitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJs [29]. Solid lines areleast-squares fitting to C · exp(α · tMgO), where C and α are fittingparameters.

real parts of the complex wave vectors (kr1, kr

2) are essential forquantum-interference-induced oscillation phenomena. Whenk‖ · �z > 0.59 (�z is the interlayer spacing of MgO(0 0 1) ),kr

1 �= kr2 (�= 0) and κ1 = κ2 = κ [13]. Tunnelling transmittance

T can then be expressed as

T = | exp(ik1 · tMgO) + exp(ik2 · tMgO)|2= 2 exp(−2κ · tMgO) · {1 + cos((kr

1 − kr2) · tMgO)}. (3)

The tunnelling transmittance for a given k‖ thus oscillates asa function of tMgO with a period proportional to 2π/(kr

1 −kr

2). This transmittance oscillation could be the origin ofthe observed oscillation of the MR-versus-tMgO curve. Onefundamental question is whether the oscillation in the MRratio (≡ (RAP − RP)/ RP) is due to the oscillation of thetunnelling resistance in the P state (RP) or of that in theAP state (RAP). RAP will theoretically oscillate becausetunnelling in the AP state dominates at certain points (hotspots) in the two-dimensional Brillouin zone (see figure 9(c))where kr

1 �= kr2(�= 0). Oscillation of RP will theoretically be

negligibly small because tunnelling in the P state dominates atthe � point (k‖ = 0 point) in the two-dimensional Brillouinzone (see figure 9(a)), where the complex wave vectorshave no real part (kr

1 = kr2 = 0). Although there are

also hot spots in the minority-spin channel of the P state(see figure 9(b)), the contribution of hot-spot tunnelling isnegligibly small compared with the contribution of the �-pointtunnelling [13, 14].

A small oscillatory component should be superposedon the exponential tMgO-dependence of the RA product infigure 15. The RA–tMgO relationship for the P and AP states infigure 15 was fitted to an exponential function, C ·exp(α ·tMgO),using the least-squares method. Here C and α are fittingparameters. By removing the exponential tMgO-dependencefrom the RA-tMgO relationship (i.e by dividing RA by exp(α ·tMgO)), we extracted an oscillatory component in the RA-tMgO

relationship. Oscillatory components in RP and RAP (i.e. thetMgO-dependences of RA/exp(α · tMgO) for the P and AP states)

Figure 16. For epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJs at20 K, the relations between tMgO and (a) MR ratio, (b) theoscillatory component of tunnel resistance in the P state(RP / exp(α · tMgO)) and (c) the oscillatory component of tunnelresistance in the AP state [RAP/ exp(α · tMgO)] [29]. Dashed greenlines are background curves fitted to quadratic functions. (a), (b)Solid blue lines are least-squares fittings to a single short-period(3.17 Å) oscillation plus the background. (a), (c) Solid red lines aresuperpositions of short- and long-period oscillations plus thebackground. The period of the short-period oscillation (dotted blackline in (c)) is 3.17 Å (fixed parameter). The period of thelong-period oscillation (broken black line in (c)) is 9.92 Å.

are shown in figures 16(b) and (c). As seen in figure 16(b), RP

oscillates as a function of tMgO. This oscillation is fitted wellwith a single-period oscillation function with the same periodas that of the MR-versus-tMgO oscillation. The phase φ of theoscillation in RP (figure 16(b)) is shifted by 212◦ ± 4◦ withrespect to that of the MR oscillation (figure 16(a)). If the MRoscillation were due solely to the RP oscillation, phase shift�φ between the oscillations in the MR ratio and RP wouldbe exactly 180◦ because the MR ratio ≡ (RAP − RP)/RP =RAP/RP − 1. The observed phase shift of 212◦ indicates that

R344

Topical Review

Figure 17. (a) Second-derivative tunnelling spectrum (|d2I/dV 2| as a function of bias V ) at 6 K in the AP state for an epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ. (Adapted from [33].) (b) Band dispersion of bcc Fe in the [0 0 1] (�–H) direction. Black and grey linesrespectively represent majority-spin and minority-spin bands.

RAP also has an oscillatory component with the same λ and adifferent φ. The oscillatory component of the RAP-versus-tMgO

relation shown in figure 16(c) is complex and cannot be fittedby a single-period oscillation function. We therefore tried tofit that oscillatory component by superposing two kinds ofoscillations with different periods, one of which was fixedto 3.17 Å (the period of the MR and RP oscillations). Asshown in figure 16(c), the oscillatory component of the RAP-versus-tMgO relation is fitted well with a superposition of twooscillations: a short-period oscillation with λ = 3.17 Å (fixed)and a long-period oscillation with λ = 9.9 Å. Phase φ of theshort-period oscillation in RAP is shifted by −61◦ ± 6◦ withrespect to φ of the TMR oscillation. As mentioned above,there is a discrepancy between the observed MR ratios andthe fitting curve (solid blue line in figure 16(a)). The originof this discrepancy seems to be the long-period oscillation inRAP. As shown by the solid red line in figure 16(a), the TMRoscillation is fitted almost exactly by using both the short-period (3.17 Å) and long-period (9.9 Å) oscillations. Thus,RAP is certainly concluded to have the short- and long-periodoscillatory components.

The experimental results, surprising because anyoscillation in RP had been expected to be negligibly small,provide important clues to the mechanism of the oscillatoryTMR effect. A possible explanation is that the actualcontribution of hot-spot tunnelling in the P state is muchgreater than theoretically expected. The short- and long-periodoscillations in RAP might indicate that there are two kinds ofhot spots at different k‖ points in the AP state. It should benoted that first-principle calculations of hot-spot tunnellingare a very delicate matter and are sensitive to such variousfactors as the interface lattice parameters and the calculationmethods. Clarifying the mechanism of TMR oscillation willrequire further theoretical and experimental studies.

4.2. Spin-dependent tunnelling spectroscopy

Tunnelling spectroscopy is a bias-voltage dependence ofconductivity in tunnel junctions. According to a simplemodel [30], the derivative of an I–V curve (the differentialconductivity) is expressed as

dI/dV = (2πe2/h)|t |2D1(EF) D2(EF + eV ), (4)

where |t |2 is the tunnel probability, D1 and D2 are the densitiesof states of the two electrodes and V is the bias voltage. Bychanging the bias voltage V , we can observe the features ofthe unoccupied density of states in electrode 2. The tunnellingspectrum of a MTJ is spin-dependent, that is, it depends onthe magnetic alignment of the two electrodes because themajority-spin and minority-spin densities of states are different(see figure 1). The second-derivative conductance spectrum(d2I/dV 2 versus V ) is more suitable for measuring a detailedDOS structure. Tunnelling spectra (dI/dV versus V andd2I/dV 2 versus V ) have been experimentally studied forMTJs with an amorphous Al–O barrier [31, 32]. Althoughextrinsic structures due to magnon scatterings and phononscatterings were observed in the spin-dependent tunnellingspectra, intrinsic features due to the densities of states ofelectrodes have never been observed in Al–O-based MTJs,which has been an open question for the last decade. Thefeatureless tunnelling spectra might be due to a diffusivetunnelling process in the Al–O-based MTJs.

We measured spin-dependent tunnelling spectra inepitaxial Fe/MgO/Fe in order to observe DOS features inthe spectra [33]. If the tunnelling process in the crystallineMgO(0 0 1) barrier is actually coherent, as illustrated infigure 5(b), the spin-dependent tunnelling spectra shouldreflect the band structure of the Fe electrode. A second-derivative conductance spectrum (d2I/dV 2 versus V ) for theAP state is shown in figure 17(a). Two sets of large peaks areevident: one at ±1 V (black arrows) and the other at ±50 mV(white arrows). The peaks at ±50 mV are considered to bemagnon scattering peaks like those for the AP state in Al–O-based MTJs [31]. The peaks at ±1 V should be due to the bandstructure of electrodes because neither phonon scattering normagnon scattering can take place at such a high-energy level.These peaks are thought to be related to the bottom edge of theminority-spin Fe �1 band, VB, which is about 1.3 eV above EF

(see figure 17(b)), in the following way. When bias voltage V islower than VB (figure 18(a)), no significant tunnelling currentflows in the AP state. When V > VB, a large tunnelling currentflows in the AP state because majority-spin �1 electrons cantunnel into the minority-spin �1 band in the other electrode.Therefore, a peak should be observed at |V | = VB as shownin figure 18(c). The large peaks observed at ±1 V are thusthought to originate from the DOS of the minority-spin Fe �1

R345

Topical Review

Figure 18. (a),(b) Tunnelling process of �1 electrons in epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ. VB denotes the bottom edge ofthe minority-spin �1 band. (a) Bias V < VB and (b) bias V > VB.(c) Expected tunnelling spectra.

band even though the peak position is slightly lower than VB.This small discrepancy could be due to the proximity effect ofthe MgO layer.

4.3. An interlayer exchange coupling (IEC) mediated bytunnelling electrons

The epitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ is also amodel system used in studying IEC between two ferromagnetic(FM) layers separated by an insulating NM spacer. IEC in aFM/NM/FM structure with a metallic NM spacer has beenstudied extensively and is well known to show oscillationsas a function of the spacer thickness [34]. IEC for ametallic spacer is induced by conduction electrons at EF.Although similar IEC is also expected in a FM/NM/FMstructure with an insulating NM spacer (i.e. a MTJ) [35],such intrinsic IEC has not been observed in MTJs with anamorphous Al–O barrier. Faure-Vincent et al were the first toobserve intrinsic antiferromagnetic IEC, on which an extrinsicferromagnetic magnetostatic coupling was superposed, in anepitaxial Fe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) MTJ structure [36].We recently obtained a more refined experimental resultexhibiting little extrinsic magnetostatic coupling (figure 19)[37]. Antiferromagnetic coupling is seen at tMgO < 0.8 nm.With increasing tMgO the IEC changes sign at tMgO =0.8 nm and then gradually approaches zero. Because of theatomically flat barrier/electrode interfaces (see figure 11), noextrinsic magnetostatic coupling was observed. The IEC fora MgO(0 0 1) spacer is apparently mediated by spin-polarizedtunnelling electrons.

5. CoFeB/MgO/CoFeB MTJs for device applications

5.1. MTJ structure for practical applications

As explained in section 3, MTJs with a single-crystalMgO(0 0 1) barrier or a textured MgO(0 0 1) barrier exhibitthe giant TMR effect at RT. This is a desirable propertyfor spintronic device applications such as MRAM and theread head of a HDD. These MTJ structures, however, are

Figure 19. Interlayer exchange coupling in epitaxialFe(0 0 1)/MgO(0 0 1)/Fe(0 0 1) at room temperature. (Adaptedfrom [37].)

Figure 20. Cross-sectional structure of a MTJ for practicalapplications.

not applicable to practical devices because of the followingproblem. MTJs for practical applications need to have thestacking structure shown in figure 20. That is, they need tohave a seed layer, an antiferromagnetic (AF) layer, a syntheticferrimagnetic structure (pinned layer), a tunnel barrier anda ferromagnetic layer (free layer). Ir–Mn or Pt–Mn is usedas the antiferromagnetic layer for exchange-biasing. The topferromagnetic layer acts as the free layer of a spin-valve. Thesynthetic ferrimagnetic (SyF) structure, which consists of anantiferromagnetically coupled FM/NM/FM trilayer such asCo–Fe/Ru/Co–Fe, is exchange-biased by the AF layer and actsas the pinned layer of a spin-valve. This type of pinned layerstructure is indispensable for device applications because of itsrobust exchange-bias and small stray magnetic field acting onthe top free layer. A reliable AF/SyF pinned layer is based on afcc structure with (1 1 1)-orientation. A fundamental problemwith this structure is that a NaCl-type MgO(0 0 1) barrier(with 4-fold in-plane crystallographic symmetry) cannot begrown on the fcc(1 1 1)-oriented AF/SyF structure (with 3-fold in-plane symmetry) because of the mismatch in structuralsymmetry. In principle we could solve this problem bydeveloping a new pinned layer structure having a 4-fold in-plane symmetry. This solution, however, is not acceptable tothe electronics industry because more than a decade has beenspent developing the reliable pinned layer structure. It should

R346

Topical Review

Figure 21. (a) A sputtering deposition machine for mass-manufacturing of HDD read heads and MRAM. (b) φ8 inch thermally oxidized Siwafer, on which a CoFeB/MgO/CoFeB MTJ film was deposited.

Figure 22. Cross-sectional TEM images of a CoFeB/MgO/CoFeBMTJ with a synthetic ferrimagnetic (SyF) pinned layer and anantiferromagnetic (AF) layer for exchange-biasing underneath theMTJ part. (a) is a magnified view of part of (b). (Reprintedfrom [38].)

be noted that the reliability of practical devices is determinedby the reliability of the pinned layer.

5.2. Giant TMR effect in CoFeB/MgO/CoFeB MTJs

To solve the above-mentioned problem concerning thecrystal growth, we developed a new MTJ structure,CoFeB/MgO/CoFeB, by using sputtering deposition [38]. Inthis study, we used a Canon ANELVA C-7100 (figure 21(a)),which is a standard sputtering deposition machine for the massproductions of HDD read heads and MRAM. We depositedCoFeB/MgO/CoFeB MTJs on φ8 inch thermally oxidized Siwafers (figure 21(b)). Cross-sectional TEM images of the MTJare shown in figure 22. As seen at the higher magnification(figure 22(a)), the bottom and top CoFeB electrode layers havean amorphous structure in an as-grown state. Surprisingly,the MgO barrier layer grown on the amorphous CoFeB is(0 0 1)-oriented polycrystalline. Because the bottom CoFeBelectrode is amorphous, this CoFeB/MgO/CoFeB MTJ can begrown on any kind of underlayers. As shown in figure 22(b),for practical applications the CoFeB/MgO/CoFeB MTJ can begrown on the standard AF/SyF pinned layer structure. We were

Figure 23. Magnetoresistance curves at RT and 20 K at a biasvoltage of 10 mV for CoFeB/MgO/CoFeB MTJ. (Reprintedfrom [38].)

able to grow the MgO(0 0 1) barrier with 4-fold symmetry onthe practical pinned layer with 3-fold symmetry by insertingthe amorphous CoFeB bottom electrode layer between them.After annealing at 360 ◦C, the CoFeB/MgO/CoFeB MTJexhibited a MR ratio of 230% at RT (figure 23; ④ in figure 13).MR ratios of up to 500% have been obtained at RT inCoFeB/MgO/CoFeB MTJs [39]. The CoFeB/MgO/CoFeBMTJs can be fabricated by sputtering deposition at RT followedby ex situ annealing. Because this fabrication processis highly compatible with high-throughput mass-productionprocesses, current research-and-development efforts devotedto the production of spintronic devices such as MRAM andHDD read heads are based on this MTJ structure. Theseapplications are explained in section 6.

5.3. Growing a textured MgO(0 0 1) barrier on anamorphous CoFeB electrode

The textured growth of an ultrathin MgO(0 0 1) layer on anamorphous CoFeB is very unusual because room-temperaturedeposition of MgO on other amorphous underlayers usuallyresults in an amorphous MgO in the initial growth stage. Forexample, when MgO is grown on amorphous SiO2, the first

R347

Topical Review

Figure 24. The growth process for a CoFeB/MgO/CoFeB MTJ film.

Figure 25. Schematic illustration of the structure of a CoFeB/MgO/CoFeB MTJ in the as-grown state and after annealing above 250 ◦C. (a)Cap layer is Ta or Ru. (b) Cap layer is Ni0.8Fe0.2 (permalloy).

3–4 nm of the MgO becomes amorphous. When MgO is grownon amorphous CoFeB, on the other hand, a polycrystallineMgO(0 0 1) layer is obtained at MgO thicknesses (tMgO) below2 nm (see figure 22(a)). To investigate the growth mechanismin detail, we made in situ reflective high-energy electrondiffraction (RHEED) observations [40]. The observed growthprocess is shown schematically in figure 24. In the initialgrowth stage (tMgO < 1.0 nm), the MgO layer is amorphous.When tMgO exceeds 1.0 nm, MgO begins to crystallize in(0 0 1)-oriented texture. After the growth of the top CoFeBelectrode layer, the MTJ structure in an as-grown state shouldbe amorphous CoFeB/amorphous MgO(1.0 nm)/texturedMgO(0 0 1)/amorphous CoFeB (figure 24(c)). Accordingto the cross-sectional TEM image in figure 22(a),however, the MTJ structure is amorphous CoFeB/textured

MgO(0 0 1)/amorphous CoFeB (figure 24(e)). Crystallizationof the 1.0 nm thick amorphous MgO (figure 24(d)) took placeat an unintentionally elevated temperature either during thedeposition process or during the preparation of the TEMsample.

5.4. Crystallization of amorphous CoFeB electrodes duringannealing

As shown in figures 5(b) and 7, 4-fold crystallographicsymmetry in both the MgO barrier and the electrodes isessential for the coherent tunnelling of �1 states. AmorphousCoFeB electrodes are therefore not expected to result in agiant TMR effect. We observed, however, that the amorphousCoFeB electrode layers adjacent to the MgO(0 0 1) barrier

R348

Topical Review

Figure 26. MR ratios at RT for CoFeB/MgO/CoFeB MTJs withvarious cap-layer materials. (Adapted from [42].)

layer crystallized in the bcc(0 0 1) structure during annealingabove 250 ◦C (figure 25(a)) [40, 41]. It should be notedthat a Co60Fe20B20 layer adjacent to a MgO(0 0 1) layercrystallizes in the bcc(0 0 1) structure even if the stablestructure of Co60Fe20B20 is not bcc but fcc [40]. This clearlyindicates that the MgO(0 0 1) layer acts as a template forcrystallizing the amorphous CoFeB layers because of the goodlattice matching between MgO(0 0 1) and bcc CoFeB(0 0 1).This type of crystallization process is known as solid phaseepitaxy. If the crystal grains and grain boundaries in thetextured MgO(0 0 1) barrier do not have a significant influenceon the transport properties, the giant TMR effect observedin CoFeB/MgO/CoFeB MTJs can be interpreted within theframework of the theory for epitaxial MTJs because themicroscopic structure of bcc CoFeB(0 0 1)/MgO(0 0 1)/bccCoFeB(0 0 1) MTJs is basically the same as that of epitaxialMTJs. (The influence of the grain boundaries is discussed insection 5.5.)

A cap layer deposited on the top CoFeB electrodein CoFeB/MgO/CoFeB MTJs was found to have a stronginfluence on the TMR effect in some cases. In standardCoFeB/MgO/CoFeB MTJs, a Ta or Ru cap layer is usuallyused (figure 25(a)). The bottom CoFeB electrode is depositedon a Ru layer (the spacer layer of the SyF structure). In thisstandard structure, MR ratios of over 200% are obtained atRT. Tsunekawa et al deposited various cap-layer materialson CoFeB/MgO/CoFeB MTJs and found that the MR ratiosignificantly depends on the cap-layer material (figure 26) [42],probably because the cap layer can influence the crystallizationof the top CoFeB electrode. A Ni0.8Fe0.2 (permalloy) caplayer, for example, grown on amorphous CoFeB has atextured fcc(1 1 1) structure (figure 25(b)). When this structureis annealed, CoFeB crystallization from the interface withthe Ni0.8Fe0.2 layer takes place at about 200 ◦C before thecrystallization from the MgO interface takes place at about250 ◦C. Consequently, the CoFeB layer crystallizes in atextured fcc(1 1 1) structure (figure 25(b)). Because the CoFeBelectrode does not have a bcc(0 0 1) structure in this case,coherent tunnelling of �1 electrons does not occur. This resultsin a significant decrease in the MR ratio. Thus, a correct choiceof cap-layer material is necessary for obtaining giant MR ratiosin CoFeB/MgO/CoFeB MTJs.

5.5. Influence of grain boundaries in a textured MgO(0 0 1)barrier on tunnelling properties

Because the textured MgO(0 0 1) barrier in CoFeB/MgO/CoFeBMTJs is polycrystalline, there are crystal grains and grainboundaries in the MgO barrier layer. One might think thatthese grain boundaries would cause a leakage current reduc-ing both the tunnelling resistance and the MR ratio. It hasbeen reported, however, that the RA products and MR ratiosof CoFeB/MgO/CoFeB MTJs with a textured MgO(0 0 1) bar-rier are similar to those of epitaxial MTJs with a single-crystalMgO(0 0 1) barrier [24, 28, 38, 39]. This has been a very puz-zling issue, and it has been speculated that grain boundariesin a textured MgO(0 0 1) barrier do not have a strong influ-ence on the tunnelling properties. To clarify the effect of grainboundaries, we used scanning tunnelling microscopy (STM) tomake local tunnelling conductance measurements in a texturedMgO(0 0 1) barrier grown on an amorphous CoFeB layer [43].

A typical STM image of a textured MgO(0 0 1) barrieron an amorphous CoFeB is shown in figure 27(a). Grainboundaries and MgO crystal grains 2–5 nm in diameter areclearly evident, and this grain size is consistent with thecoherence length of RHEED images [40]. A line-scanprofile (figure 27(b)) revealed that grain boundaries are dipswith an average depth of 0.2 nm, which corresponds to onemonolayer (ML) of MgO(0 0 1). We used scanning tunnellingspectroscopy [43] to investigate the local tunnelling, andfigure 27(c) shows typical local tunnelling spectra measured atthe centre of a grain (red lines) and at grain boundaries (bluelines). The local I–V curves showed strong nonlinearity andnearly the same shape. This means that not only the interior ofthe grains but also grain boundaries act as a nearly perfecttunnelling barrier. The tunnelling current flows uniformlydespite the existence of grain boundaries. This is consistentwith the RA products and MR ratios of CoFeB/MgO/CoFeBMTJs being similar to those of epitaxial MgO-based MTJs.Although why the tunnelling current flows uniformly in atextured MgO(0 0 1) barrier is unclear, that it does is veryfavourable for device applications.

6. Device applications of MgO-based MTJs

From the viewpoint of device applications, the history ofspintronics is a history of magnetoresistance at RT (figure 28).GMR spin-valve devices have MR ratios of 5–15% at RT andhave been used in the read heads of HDDs. Al–O-based MTJshave MR ratios of 20–70% at RT and have been used not onlyin HDD read heads but also in MRAM cells. Giant TMRMgO-based MTJs have MR ratios of 200–500% at RT and areexpected to be used in various spintronic devices such as HDDread heads, spin-transfer MRAM cells and novel microwavedevices.

6.1. Read heads of HDDs

As explained in section 5, the CoFeB/MgO/CoFeB MTJsshowing the giant TMR effect are compatible with the mass-manufacturing process for spintronics devices because theycan be fabricated on the practical pinned layer structureby sputtering deposition at RT followed by post-annealing.Besides requiring giant MR ratios and manufacturing

R349

Topical Review

Figure 27. (a) In situ scanning tunnelling microscopy (STM) image of a textured MgO(0 0 1) barrier grown on amorphous CoFeB. (b)Line-scan profile of the MgO surface. (c) Typical local tunnelling spectra measured at the centre of a grain (red lines) and at the grainboundary (blue lines). (Reprinted from [43].)

Figure 28. History and future prospects of magnetoresistive effects, MR ratios at room temperature and applications of MR effects inspintronic devices.

compatibility, industrial applications require MTJs that haveMR ratios that depend little on bias voltage, that are robustand have high break-down voltages, that can be manufactureduniformly and reproducibly, and that have appropriateresistance-area (RA) products. CoFeB/MgO/CoFeB MTJssatisfy these requirements

Because impedance matching in an electronic circuit isindispensable for a high-speed operation of an electronicdevice, the RA product of MTJs should be adjusted to satisfythe impedance-matching condition. MRAM applicationsrequire a RA in the range from 50 � µm2 to 10 k� µm2,depending on the lateral MTJ size (i.e. areal density ofMRAM). In this RA range, MR ratios of over 200% at RTcan be easily obtained using MgO-based MTJs. On the otherhand, the read head of a high-density HDD requires a verylow RA product. MTJs with an amorphous Al–O or Ti–Obarrier are currently used in TMR read heads for HDDs withareal recording densities of 100–130 Gbit/inch2 [44]. TheseMTJs have low RA products (2–3 � µm2) and MR ratios

of 20–30% at RT. Although these properties are enough forrecording densities of 100–130 Gbit/inch2, even lower RAproducts and much higher MR ratios are needed for recordingdensities above 200 Gbit/inch2. For example, a RA productbelow 1 � µm2 and a MR ratio of above 50% are requiredfor areal recording densities above 500 Gbit/inch2 (figure 29).Such a low RA product and a high MR ratio have never beenobtained in a conventional MTJ with an amorphous Al–O orTi–O barrier (figure 29). A current perpendicular to plane(CPP) GMR device, which is one of the candidates for thenext-generation HDD read head, has an ultralow RA product(below 1 � µm2), but the MR ratio of a CPP GMR device istoo low (<10% for a practical spin-valve structure) for a deviceused as a HDD read head (figure 29).

To reduce the RA product, we made CoFeB/MgO/CoFeBMTJs with an ultrathin textured MgO(0 0 1) barrier (tMgO =1.0 nm, which corresponds to only 4–5 ML) [45,46]. Because1.0 nm is the critical thickness for the crystallization of MgOgrown on an amorphous CoFeB (see section 5.2), careful

R350

Topical Review

Figure 29. MR ratio at RT versus resistance-area (RA) product.Open circles are values for CoFeB/MgO/CoFeB MTJs. (Adaptedfrom [46].) Light grey and dark grey areas are the zones required forHDDs with recording densities above 250 and 500 Gbit/inch2.

Figure 30. Cross-sectional TEM images for an ultralow-resistanceCoFeB/MgO/CoFeB MTJ with a synthetic ferromagnetic (SyF)pinned layer exchange-biased by an antiferromagnetic Pt–Mn layer.(Adopted from [46].)

optimizations of the growth conditions are necessary forgrowing a 1.0 nm thick textured MgO(0 0 1) barrier. Forexample, the textured growth of MgO layers was found tobe sensitive to the base pressure of the sputtering chamber.Residual H2O molecules in the chamber were found to degradethe crystalline orientation of MgO(0 0 1), reducing the MRratio. By carefully removing residual H2O molecules by usinga Ta getter technique, we were able to grow a highly textured1.0 nm thick MgO(0 0 1) barrier (figure 30) [46]. The removalof residual H2O molecules was also found to be effective forpreventing oxidation of the CoFeB/MgO interface [47]. TheMR ratios for these CoFeB/MgO/CoFeB MTJs are shownby the open circles in figure 29. We obtained both ultralowRA products (0.4–1 � µm2) and high MR ratios (>50%),satisfying the requirements for areal recording densities wellabove 500 Gbit/inch2.

By using ultralow-resistance MgO-based MTJs, FujitsuCorp. developed a TMR read head for ultrahigh-densityHDDs. A cross-sectional TEM image of it is shown in

Figure 31. Cross-sectional TEM image of MgO-TMR read head forHDD with recording density of 250 Gbit/inch2. (Courtesy of FujitsuCorporation.)

Figure 32. (a) Magnetoresistance curve (R–H loop) and(b) spin-transfer switching (STS) curve (R–I loop) at roomtemperature in a CoFeB/MgO/CoFeB MTJ with a lateral size of70 × 160 nm. (Adapted from [52].)

figure 31. MgO TMR heads have already been applied toHDDs with recording densities above 250 Gbit/inch2 and areexpected to later be applicable to recording densities up toabout 1 Tbit/inch2. The MgO TMR head is thus going to bethe mainstream technology for HDD read heads for at least thenext five years.

6.2. Spin-transfer MRAM

The giant TMR effect in MgO-based MTJs is also usefulin developing MRAM. In conventional MRAM, the writingprocess (i.e. magnetization reversal of a free layer) uses amagnetic field generated by pulse currents, and the read-out process uses a resistance change between parallel andantiparallel magnetic states (i.e. the TMR effect). The giantTMR effect enables high-speed read-out because a giantMR ratio yields a very high output signal [48]. In theconventional MRAM, however, the writing pulse currentsincrease when the lateral size of MTJs is reduced, whichmakes it difficult to develop Gbit-scale high-density MRAM.In a new type of MRAM called spin-transfer MRAM orspin RAM, on the other hand, the writing process usesthe magnetization switching induced by spin-transfer torque.This phenomenon, called spin-transfer switching (STS) [49],is especially important in developing high-density MRAMbecause the writing pulse current flowing through the MTJcan be reduced by reducing the lateral size of the MTJ. STSwas experimentally demonstrated first in CPP GMR devices

R351

Topical Review

Figure 33. Cross-sectional TEM images of a 4 kbit spin-transfer MRAM using CoFeB/MgO/CoFeB MTJs (courtesy of Sony Corporation).(Adapted from [56].)

[50] and later in Al–O-based MTJs [51]. STS in MgO-basedMTJs, which is especially important for MRAM, has beendemonstrated using CoFeB/MgO/CoFeB MTJs [52–56]. Anexample of STS is shown in figure 32. Switching betweenparallel and antiparallel magnetic states is induced not only byapplying a magnetic field (figure 32(a)) but also by sendinga pulse current through an MTJ (figure 32(b)). Prototypespin-transfer MRAM cells based on the giant TMR effectand STS in MgO-based MTJs have been developed [56, 57].The 4 kbit spin-transfer MRAM developed by Sony Corp., forexample, (shown in figure 33) provides reliable read-out andwrite operations [56]. At the present stage, the intrinsic criticalcurrent density Jc0, or switching pulse-current density whenthe pulse duration is 1 ns is about 2×106 A/cm2 [55,56] and isnot small enough for high-density MRAM. If Jc0 is reduced toabout 5 × 105 A/cm2, it will be possible to develop Gbit-scalespin-transfer MRAM.

6.3. Novel microwave applications

MgO-based MTJs are also potentially useful for microwavedevice applications. We demonstrated that a dc voltage isgenerated between the two electrodes when an ac current witha microwave frequency flows through a CoFeB/MgO/CoFeBMTJ with a lateral size of 100 × 200 nm [58]. Thisphenomenon, called the spin-torque diode effect, resultsfrom a combination of the giant TMR effect and a resonantprecession of free-layer magnetic moment induced by spin-transfer torque. In other words, the spin-torque diode effect isspin-torque-induced ferromagnetic resonance (FMR) detectedelectrically by using magnetoresistance. MgO-based MTJscan thus act as microwave detectors.

An inverse of the spin-torque diode effect is the microwaveemission that occurs when a dc current flows through a 100 nmscale magnetoresistive device. A spin-transfer torque acting onthe free-layer magnetic moment can induce a steady precessionof the free-layer moment at a FMR frequency. The steadyprecession of the free-layer moment in the magnetoresistivedevice induces an ac current or ac voltage at a FMR frequency(i.e. a microwave frequency). This microwave emissionwas first demonstrated using a CPP GMR device with aMR ratio of about 1% at RT [59]. The microwave powerfrom a CPP GMR device, however, is only of nanowattorder and is too small for practical use. But because the

microwave power is theoretically proportional to the squareof the MR ratio, MgO-based MTJs with giant MR ratiosare potentially able to emit high-power microwaves. Werecently obtained microwave emission of microwatt orderfrom a 100 nm CoFeB/MgO/CoFeB MTJ with a MR ratio ofabout 100% [60]. Spin-torque-induced microwave emissionbased on MgO-based MTJs is expected to be the third majorapplication of spintronics technology (see figure 28).

7. Conclusion

In 2001, first-principle theories predicted that epitaxialFe/MgO/Fe MTJs with a crystalline MgO(0 0 1) barrier wouldhave MR ratios of over 1000% [13, 14]. The mechanismof the TMR effect in MgO-based MTJs is different fromthat in conventional MTJs with an amorphous Al–O barrier.In conventional MTJs, various Bloch states with differentsymmetries can tunnel incoherently through an amorphousAl–O barrier, which results in a reduction of tunnelling spinpolarization and thus in MR ratios below about 70% at RT.In MgO-based MTJs, on the other hand, only Bloch stateswith �1 symmetry dominantly tunnel through a crystallineMgO(0 0 1) barrier because the barrier acts as a symmetry filter.Because the �1 Bloch states in bcc ferromagnetic metals andalloys based on Fe and Co (e.g. Fe, Fe–Co, CoFeB and someHeusler alloys) are at EF fully spin-polarized in the [0 0 1]direction, the giant TMR effect is expected when a crystallineMgO(0 0 1) barrier is combined with these ferromagneticelectrode materials.

In 2004, MR ratios of up to about 200% at RT wereobtained experimentally by using fully epitaxial MTJs witha single-crystal MgO(0 0 1) barrier [24, 25] and MTJs with ahighly oriented polycrystalline (textured) MgO(0 0 1) barrier[27]. MR ratios of over 400% at RT have been observed inepitaxial MgO-based MTJs [28].

Epitaxial MTJs with a single-crystal MgO(0 0 1) barrierare a model system for studying the physics of spin-dependenttunnelling because of their well-defined structure. In additionto the giant TMR effect, epitaxial MgO-based MTJs exhibitother interesting phenomena not exhibited by Al–O-basedMTJs. For example, the MR ratio of an epitaxial Fe/MgO/FeMTJ oscillates as a function of the MgO barrier thicknesstMgO [25, 29]. While the tunnelling resistance for theparallel magnetic state shows a single-period (short-period)

R352

Topical Review

oscillation as a function of tMgO, the tunnelling resistance forthe antiparallel magnetic state shows short- and long-periodoscillations as a function of tMgO [29]. As a result, the tMgO-dependence of the MR ratio is expressed as a superpositionof the short- and long-period oscillations. Although thedetailed mechanism of the oscillatory TMR effect is not clear,the oscillation is thought to be related to the coherency oftunnelling electrons. The epitaxial Fe/MgO/Fe MTJs alsoexhibit complex spin-dependent tunnelling spectra due to theband structure of Fe [33] and IEC mediated by tunnellingelectrons [36, 37]. Clarifying the mechanisms of thesephenomena will lead to a deeper understanding of the physicsof spin-dependent tunnelling.

Epitaxial and textured MgO-based MTJs are not suitablefor use in MRAM or HDD read heads because they cannotbe grown on the practical synthetic ferrimagnetic (SyF)pinned layer, which has 3-fold in-plane crystallographicsymmetry. We therefore developed a novel MTJ structure:CoFeB/MgO/CoFeB [38]. We found that a highly texturedMgO(0 0 1) barrier layer can be grown on an amorphousCoFeB bottom electrode layer by sputtering deposition atRT. A CoFeB/MgO/CoFeB MTJ can be grown on any kindof underlayer, including the practical SyF pinned layer.Annealing CoFeB/MgO/CoFeB MTJs above 250 ◦C causesthe amorphous CoFeB electrode layers to crystallize in thebcc(0 0 1) structure and results in MR ratios of above 200%at RT [40]. Because of the good lattice matching betweenMgO(0 0 1) and bcc CoFeB(0 0 1), the MgO(0 0 1) layeracts as a template for crystallizing CoFeB in the bcc(0 0 1)structure. Because the annealed MTJ structure is texturedbcc CoFeB(0 0 1)/MgO(0 0 1)/bcc CoFeB(0 0 1), the observedgiant TMR effect can be explained within the framework ofthe theories for epitaxial MgO-based MTJs. Although theMgO barrier in CoFeB/MgO/CoFeB MTJs is a polycrystallinebarrier with grain boundaries, in situ STM study revealedthat a tunnelling current flows uniformly through the texturedMgO(0 0 1) barrier without leakage at the grain boundaries[43]. This is consistent with the fact that the RA productsand MR ratios of CoFeB/MgO/CoFeB MTJs are similar tothose of epitaxial MgO-based MTJs. This is very favourablefor device applications because a textured MgO(0 0 1) barrieris as uniform and reliable as a single-crystal MgO(0 0 1)barrier.

The giant TMR effect in MgO-based MTJs, especiallyin CoFeB/MgO/CoFeB MTJs, is of great importance fornext-generation spintronic devices. Ultralow-resistanceCoFeB/MgO/CoFeB MTJs with an ultrathin MgO(0 0 1)barrier were developed for use in HDD read heads [45,46]. A MgO TMR read head for HDDs with recordingdensities above 250 Gbit/inch2 has already been developed andcommercialized, and future MgO TMR heads are expected toprovide recording densities up to 1 Tbit/inch2. A prototypespin-transfer MRAM based on CoFeB/MgO/CoFeB MTJs hasbeen developed and shown to provide reliable read-out andwrite operations [56,57]. The giant TMR effect in MgO-basedMTJs is also useful in developing novel microwave detectorsand oscillators. In these applications, output performance isroughly proportional to the MR ratio at RT. The giant TMReffect in the MTJs is therefore expected not only to extend theapplications of existing devices but also to help realize novelspintronic applications.

Acknowledgments

This work has been done in collaboration with A Fukushima,T Nagahama, H Kubota, A A Tulapurkar, K Ando(AIST), K Tsunekawa, H Maehara, Y Nagamine, M Nagai,S Yamagata, Y S Choi, N Watanabe (Canon ANELVACorporation), R Matsumoto, M Mizuguchi, A Deac andY Suzuki (Osaka University). It was partly supported bythe New Energy and Industrial Technology DevelopmentOrganization (NEDO) of Japan. We are grateful to theseveral colleagues who gave kindly allowed us to modify theirpreviously published figures.

References

[1] Julliere M 1975 Phys. Lett. 54A 225[2] Baibich M N et al 1988 Phys. Rev. Lett. 61 2472[3] Binasch G, Grunberg P, Saurenbach F and Zinn W 1989 Phys.

Rev. B 39 4828[4] Miyazaki T and Tezuka N 1995 J. Magn. Magn. Mater. 139

L231[5] Moodera J S, Kinder L R, Wong T M and Meservey R 1995

Phys. Rev. Lett. 74 3273[6] Wang D, Nordman C, Daughton J, Qian Z and Fink J 2004

IEEE Trans. Magn. 40 2269[7] Meservey R and Tedrow P M 1994 Phys. Rep. 238 173[8] Parkin S S P et al 2003 Proc. IEEE 91 661[9] Bowen M et al 2003 Appl. Phys. Lett. 82 233

[10] Sakuraba Y et al 2006 Appl. Phys. Lett. 88 192508[11] Yuasa S, Nagahama T and Suzuki Y 2002 Science 297 234[12] Nagahama T, Yuasa S, Tamura E and Suzuki Y 2005 Phys.

Rev. Lett. 95 086602[13] Butler W H, Zhang X-G, Schulthess T C and MacLaren J M

2001 Phys. Rev. B 63 054416[14] Mathon J, Umersky A 2001 Phys. Rev. B 63 220403R[15] Bagayako D, Ziegler A and Callaway J 1983 Phys. Rev. B

27 7046[16] Zhang X-G and Butler W H 2004 Phys. Rev. B 70 172407[17] Mavropoulos Ph, Papanikolaou N, Dederichs P H 2000 Phys.

Rev. Lett. 85 1088[18] Velev J P et al 2005 Phys. Rev. Lett. 95 216601[19] Wulfhekel W et al 2001 Appl. Phys. Lett. 78 509[20] Bowen M et al 2001 Appl. Phys. Lett. 79 1655[21] Faure-Vincent J et al 2003 Appl. Phys. Lett. 82 4507[22] Meyerheim H L et al 2001 Phys. Rev. Lett. 87 07102[23] Zhang X-G, Butler W H and Bandyopadhyay A 2003 Phys.

Rev. B 68 092402[24] Yuasa S, Fukushima A, Nagahama T, Ando K and Suzuki Y

2004 Japan. J. Appl. Phys. 43 L588[25] Yuasa S, Nagahama T, Fukushima A, Suzuki Y, Ando K 2004

Nature Mater. 3 868[26] Miyokawa K et al 2005 Japan. J. Appl. Phys. 44 L9[27] Parkin S S P et al 2004 Nature Mater. 3 862[28] Yuasa S, Fukushima A, Kubota H, Suzuki Y and Ando K 2006

Appl. Phys. Lett. 89 042505[29] Matsumoto R, Fukushima A, Nagahama T, Suzuki Y, Ando K

and Yuasa S 2007 Appl. Phys. Lett. 90 252506[30] Wiesengdanger R 1994 Scanning Probe Microscopy and

Spectroscopy (Cambridge: Cambridge University Press)[31] Moodera J S, Nowak J and van de Veerdonk R J M 1998 Phys.

Rev. Lett. 80 2941[32] Ando Y, Murai J, Kubota H and Miyazaki T 2000 J. Appl.

Phys. 87 5209[33] Ando Y et al 2005 Appl. Phys. Lett. 87 142502[34] Parkin S S P, More N and Roche K 1990 Phys. Rev. Lett.

64 2304[35] Bruno P 1995 Phys. Rev. B 52 411[36] Faure-Vincent J et al 2002 Phys. Rev. Lett. 89 107206[37] Katayama T et al 2006 Appl. Phys. Lett. 89 112503

R353

Topical Review

[38] Djayaprawira D D et al 2005 Appl. Phys. Lett. 86 092502[39] Lee Y M et al 2007 Appl. Phys. Lett. 90 212507[40] Yuasa S, Suzuki Y, Katayama T and Ando K 2005 Appl. Phys.

Lett. 87 242503[41] Choi Y S et al 2007 J. Appl. Phys. 101 013907[42] Tsunekawa K et al 2005 Digests of the IEEE Int. Magnetics

Conf. (Intermag, Nagoya, Japan) HP-08[43] Mizuguchi M, Suzuki Y, Nagahama T and Yuasa S 2007 Appl.

Phys. Lett. 91 012507[44] Zhu J-G and Park C 2006 Mater. Today 9 (11) 36[45] Tsunekawa T et al 2005 Appl. Phys. Lett. 87 072503[46] Nagamine Y et al 2006 Appl. Phys. Lett. 89 162507[47] Choi Y S et al 2007 Appl. Phys. Lett. 90 012505[48] Slaughter J M et al 2005 Technical Digest of IEEE Int.

Electron Devices Meeting (IEDM (Washington, DC)) 35.7

[49] Slonczewski J C 1996 J. Magn. Magn. Mater. 159 L1[50] Katine J A et al 2000 Phys. Rev. Lett. 84 3149[51] Huai Y et al 2004 Appl. Phys. Lett. 84 3118[52] Kubota H et al 2005 Japan. J. Appl. Phys. 44 L1237[53] Kubota H et al 2006 Appl. Rev. Lett. 89 032505[54] Hayakawa J et al 2005 Japan. J. Appl. Phys. 44 L1267[55] Diao Z et al 2005 Appl. Phys. Lett. 87 232502[56] Hosomi M et al 2005 Technical Digest of IEEE Int. Electron

Devices Meeting (IEDM (Washington, DC)) 19.1[57] Kawahara T et al 2007 Technical Digest of IEEE Int.

Solid-State Circuits Conf. (ISSCC, San Francisco, CA) 26.5[58] Tulapurkar A A et al 2005 Nature 438 339[59] Kiselev S I et al 2003 Nature 425 380[60] Deac A et al 2007 Abstracts of Joint MMM-Intermag Conf.

(Baltimore, MD) FD-02

R354