Main

Resistivity values ≈h/e2 indicate that the electron mean free path l is shorter than the Fermi wavelength λF, so that quantum interference becomes a dominant feature in electron diffusion, leading to Anderson localization in the absence of phase-breaking processes at low temperatures (T). The scope of this phenomenon extends beyond electronic systems—into optical and acoustic phenomena as well1,2,3—but not generic graphene, which remains metallic at liquid-helium T (refs 4, 5) and exhibits only a weak T dependence that can be explained by phonons and thermally excited carriers11. Earlier theoretical studies have suggested that Dirac electrons can evade localization for certain types of disorder3,12,13,14,15, with the extreme example being graphene subjected to a smooth Coulomb potential16,17. However, for generic disorder that involves scattering between the two graphene valleys, the localization is expected to be unavoidable3,18,19. Experiments do not show this.

In this Letter, we describe a double-layer electronic system made of two closely-spaced but electrically isolated graphene monolayers sandwiched in boron nitride. In the following, the two layers in the double layer graphene (DLG) heterostructure are referred to as the studied and control layers. At low doping nC in the control layer, the studied layer exhibits the standard behaviour with a minimum metallic conductivity of 4e2/h. However, for nC>1011 cm−2, the resistivity ρ of the studied layer diverges near the neutrality point (NP) at T<70 K. This divergence can be suppressed by a small perpendicular field B<0.1 T, which indicates that this is an interference effect rather than a gap opening. We attribute the metal–insulator transition (MIT) to the recovery of an intrinsic behaviour such that graphene exhibits Anderson localization if its ρ reaches values of ≈h/e2 per carrier type. Normally, this intrinsic MIT is obscured by charge inhomogeneity in the form of electron–hole puddles20,21,22,23,24. Within each puddle, graphene is sufficiently away from the NP and remains metallic. Then, resistivity of the percolating electron–hole system with leaking p–n boundaries16,17 assumes a value of h/e2 with little T dependence (conceptually this value has little in common with the similar value required for Anderson localization) 23,24. The control layer can screen out the fluctuating background potential and suppress electron–hole puddles, revealing the intrinsic properties at the NP. This reconciles the metallic behaviour normally observed in graphene with the localization expected for large ρ and supports the idea that the minimum conductivity that tends to assume values close to 4e2/h is due to electron–hole puddles23,24.

The studied devices were fabricated by sandwiching two graphene monolayers with thin hexagonal-BN crystals. In a multistep procedure, described in the Supplementary Information, a graphene monolayer was transferred onto a 20–30 nm thick BN crystal that was first prepared on top of an oxidized Si wafer. Then, the graphene was covered with another BN crystal (spacer), which was followed by transfer of the second graphene layer. Both layers were shaped into multiterminal devices aligned above each other and having separate electrical contacts (Fig. 1a). Individual steps were similar to those described in refs 25, 26 but the whole fabrication process involved three dry transfers and alignments, four nonconsecutive rounds of electron-beam lithography, three rounds of plasma etching and two separate metal depositions. The resulting DLG heterostructures are schematically shown in Fig. 1a (for images, see Supplementary Information). We made several such devices with channel widths of 1–2 μm. They exhibited μ of 30– 120×103 cm2 V−1 s−1 and little chemical doping. The bottom layer encapsulated in BN always had higher μ and changed little after exposure to air26 whereas the quality of the top layer gradually decayed. For this particular study, we employed three multiterminal devices with sufficiently thick BN spacers to avoid any detectable tunnel current between graphene layers (<0.1 nA). The spacers had thicknesses d≈4, 12 and 16 nm. All the devices exhibited a similar MIT behaviour, although the insulating state was much more pronounced for devices with smaller d and higher μ, as described below.

Figure 1: Electron transport in graphene–BN heterostructures.
figure 1

a, Schematic view of our heterostructure devices and measurement geometry. b,c, ρ as a function of n in the studied graphene layer for different doping nC of the control layer at two temperatures. The device has a 4 nm BN spacer.

With reference to Fig. 1a, we employed the following scheme of measurements. A voltage V t was applied between the graphene layers, and this electrically doped both of them with carriers of the opposite sign. The bottom layer could also be gated by a voltage V b applied to the Si wafer. Because of the low density of states, graphene can provide only a partial screening and, therefore, V b induced carriers in the top layer as well. This influence was weaker than on the bottom layer and depended on n in the latter. By measuring the Hall resistivity ρx y we could determine n in each of the layers (Supplementary Information). We usually fixed V tto define a nearly constant n in the top layer and swept V b to vary n in the bottom layer. Normally, we studied the higher- μ bottom layer and used the top layer as control. In this configuration, the insulating state reached higher ρ. If the studied and control layers were swapped, the behaviour remained qualitatively the same (Supplementary Information) but lower μ resulted in lower ρ of the insulating state.

Our main result is illustrated by Fig. 1, which shows two sets of standard curves ρ(n)for the studied layer at different nC. At 70 K, the control layer has little effect on the studied layer, and all the curves in Fig. 1b look no different from those observed in the standard devices4 or for graphene on BN (GBN; ref. 25). However, at low T and for high doping of the control layer (nC>1011 cm−2), graphene exhibits a radically different behaviour (Fig. 1c). In this regime, ρ at the NP acquires a strong T dependence and easily overshoots the threshold value of h/e2. To elucidate this observation, Fig. 2 shows further examples of ρ(n,T) for high and low doping of the control layer. In the case of large nC (Fig. 2a), ρ exhibits an insulating T dependence. In contrast, zero nC results in a much weaker T dependence that can be explained by thermally excited carriers (Fig. 2b). Outside a relatively narrow interval |n|≤1010 cm−2 and above 70 K, the behaviour of graphene was practically independent of nC.

Figure 2: Resistivity of the studied layer at different T for high and low doping of the control layer.
figure 2

a,b, Correspond to nC≈3×1011 cm−2 and zero nC, respectively. Here, we have chosen to plot data for d≈12 nm. For our thinnest spacer (≈4 nm), ρNPbecomes very large at low T(inset) and continuous curves ρ(n) are difficult to measure because of crosstalk nonlinearities (Supplementary Information). The inset shows the T dependence of ρNP for the device in the main figure at both nC (open and filled circles) and for the 4 nm device at nC≈5×1011 cm−2 (squares). The dashed line indicates the threshold value for Anderson localization, ρ=h/4e2.

The T dependence of the maximum resistivity at the NP, ρNP, is shown in more detail in the inset of Fig. 2b for zero and high doping of the control layer. The insulating state is more pronounced for d=4 nm but remains clear also for the 12 nm device (note the logarithmic scale). For d=4 nm and below 4 K, ρNP could reach into the MΩ range (Supplementary Information), an increase by 2–3 orders of magnitude with respect to the standard behaviour. The high- ρ regime is found to be difficult to probe because of a strong nonlinearity caused by a crosstalk between the measurement current and V t, the effect specific to DLG devices with small d (see Supplementary Information). To assure the linear response in this regime, we had to measure IV curves at every gate voltage and, to avoid these difficulties, we limited our studies mostly to T>4 K and ρNP<100 kΩ.

The influence of the adjacent layer immediately invites one to consider interlayer Coulomb interactions. Indeed, the relevant energy scale is e2/ɛ d50 meV, that is, the interactions may be a significant factor (ɛ≈5 is BN’s dielectric constant). For example, one can imagine that the interactions open an excitonic-like gap at the Dirac point. We have ruled out this possibility by magnetic field measurements. In the gapped case, B is expected to enhance the confinement and, hence, the binding energy. In contrast, our devices exhibit a pronounced negative magnetoresistance in non-quantizing B (Fig. 3). The insulating behaviour is suppressed in characteristic B*≈10 mT (Fig. 3), well below the onset of Landau quantization. Figure 3 also shows that the MIT is again confined to |n|≤1×1010 cm−2.

Figure 3: Resistivity of the studied layer in the insulating regime at various B.
figure 3

d=12 nm; nC≈3×1011 cm−2. At low T, ρ(n) exhibits pronounced ‘mesoscopic’ fluctuations (for example, the left shoulder in this figure) which develop further with decreasing T and are probably due to macroscopic charge inhomogeneity. Inset—detailed B dependence of ρNP for the case shown in the main figure.

Another interesting observation is that the insulating state has always developed at ρ>h/4e2 (Figs 13). This is seen most clearly in the inset of Fig. 2, where the curves depart from each other above the dashed line marking h/4e2. In the insulating state, ρNP is found to follow a power-law dependence 1/Tν, where ν varied from sample to sample, reaching a value close to two in the device with d≈4 nm. The characteristic T at which the insulating state started to develop can be attributed to the fact that above 70 K the concentration of excited carriers at the NP exceeded ≈1010 cm−2, beyond which no MIT could be observed even at low T.

The suppression of the MIT by non-quantizing B is a clear indication that localization plays an important role, such that B breaks down the time-reversal symmetry and destroys the interference pattern that developed due to self-intersecting trajectories1,2,3. The strong localization scenario is also consistent with the onset of the insulating state at ρNPh/4e2, which corresponds to the resistivity quantum per carrier type. However, localization in graphene cannot possibly be explained without intervalley scattering3,18,19. A tempting line of argument would be to invoke charge fluctuations in the control layer to explain its influence on the studied layer. However, this contradicts the fact that μcan increase notably at high nC, that is, graphene exhibits higher quality rather than extra scattering if the control layer is strongly doped (Supplementary Information). Moreover, the Coulomb interaction between the layers is generally expected to become less efficient with decreasing T and increasing nC (ref. 27), which is exactly opposite to what we observe. Finally, an interlayer scattering mechanism can be ruled out by the fact that any interaction potential created by carriers in the control layer and acting on the studied layer varies at distances of da (a is the lattice constant), whereas the fast components needed for intervalley scattering depend exponentially on a/d (ref. 28).

To explain the MIT, we assume a small amount of intervalley scatterers already present in our devices. They could be either some of the defects that limit μ (for example, strong adsorbates)29 or, alternatively, the intervalley scattering can arise because of the atomic-scale potential created by BN. In both cases, this can break down the symmetry between the carbon sublattices and act as a source of intervalley scattering. Because the insulating state is observed only for |n|≤1010 cm−2 and the process responsible for Anderson localization should provide a mean free path of about λF=(4π/n)1/2, we can estimate the intervalley scattering length liv as 0.1 μm.

Furthermore, B*10 mT yields a spatial scale (ϕ0/B*)1/2≈0.5 μm, which corresponds to a flux quantum ϕ0=h/e enclosed by diffusive trajectories. This scale is significantly larger than the mean free path l≤0.1 μmestimated for the relevant interval of n≤1010 cm−2 and, therefore, this justifies the use of diffusive transport concepts. Fitting the magnetoresistance curves, such as in Fig. 3, by the weak localization formulas19,30 (although mentioning that those are applicable to small rather than large changes in ρ) yields two other spatial scales. One corresponds to the onset of magnetoresistance (1 mT) and yields the phase-breaking length of a few μm at liquid-helium T, which is typical for graphene30,31. The other scale (≈0.1 μm) is given by B≈0.1 T, where the magnetoresistance saturates, before changing its sign from negative to positive. The latter scale could be due to the onset of intervalley scattering18,19,31, which agrees well with the value liv determined from the above analysis of the MIT.

The proposed scenario for the MIT can be considered routine for any high- ρ metallic system at low T, including the previously studied damaged graphene, which contains a large amount of short-range, intervalley scatterers7,8,9,10. Quality graphene has been the only known exception until now. Therefore, the question should be turned around and it should be asked why there is no MIT in the standard graphene devices or DLG at low nC and why the MIT becomes pronounced only in our ultra-high-quality graphene. The latter seemingly contradicts the very notion of Anderson localization. The puzzle has a straightforward resolution if we attribute this behaviour to the presence of electron–hole puddles4,23,24.

In graphene on SiO2, the puddles contain carriers in typical n1011 cm−2 (ref. 20). In GBN, puddles are larger and shallower21,22 but, within each puddle, n is still high enough (>1010 cm−2) to move the system away from the MIT. The resistivity of such an inhomogeneous system is then determined by inter-puddle ballistic transport with ρh/4e2 (refs 23, 24). The recovery of the MIT can be expected if n within the electron–hole puddles decreases below the localization threshold (≈1010 cm−2 in our case). Accordingly, we attribute the influence of the control layer to the fact that at high nC it screens out the background potential, making puddles shallower, as our numerical modelling shows (see Supplementary Information). Experimentally, this is also the case, as seen from Hall measurements where the transition region in ρx y (between electron- and hole-regimes) narrows at high nC (Supplementary Information). Further work is required to understand the underlying physics in detail and, especially, the mechanism of intervalley scattering and a possible role of the atomic washboard created by BN.