The high magnetic field electronic structure of bilayer graphene is enhanced by the spin, valley isospin, and an accidental orbital degeneracy, leading to a complex phase diagram of broken symmetry states. Here, we present a technique for measuring the layer-resolved charge density, from which we directly determine the valley and orbital polarization within the zero energy Landau level. Layer polarization evolves in discrete steps across 32 electric field-tuned phase transitions between states of different valley, spin, and orbital order, including previously unobserved orbitally polarized states stabilized by skew interlayer hopping. We fit our data to a model that captures both single-particle and interaction-induced anisotropies, providing a complete picture of this correlated electron system. The resulting roadmap to symmetry breaking paves the way for deterministic engineering of fractional quantum Hall states, while our layer-resolved technique is readily extendable to other two-dimensional materials where layer polarization maps to the valley or spin quantum numbers.
The single-particle energy spectrum of a two-dimensional electron system (2DES) in a large magnetic field collapses into Landau levels (LLs) containing NΦ degenerate states, with NΦ the number of magnetic flux quanta penetrating the sample. The width in energy of the LL bands is limited only by disorder, making electronic interactions effectively strong in a clean system even when their absolute scale is weak. The simplicity of the starting LL wavefunctions, combined with the high degree of control available in 2DES, make LLs a promising venue for engineering electronic ground states based on electron correlations. However, the difficulty of simulating interacting electron problems necessitates experimental input to constrain the possible ground states, particularly in the presence of internal degeneracy. The Bernal bilayer graphene (B-BLG) zero energy Landau level (ZLL) provides an extreme example of such degeneracy. In B-BLG, LLs are labeled by their electron spin σ = ↑, ↓, valley ξ = +, −, and orbital index . Electrons in valley +/− are localized near points K/K′ of the hexagonal Brillouin zone, while the index N is closely analogous to the LL-index of conventional LL systems. The energies of the LLs are approximately , where ħωc is the cyclotron energy, leading to an eight-fold nearly degenerate ZLL comprising the N = 0 and N = 1 orbitals and all possible spin and valley isospin combinations.
While direct probes of spin in 2DESs were demonstrated two decades ago1, the valley quantum number has only been probed indirectly in semiconductor quantum wells2, graphene monolayer3, 4 and bilayers5,6,7,8,9,10,11,12,13, and transition metal dichalcogenides14. In bilayer graphene, resolving the order in which the eight components fill as electrons are added is one of the key open questions, and is essential to efforts to use bilayer graphene to engineer exotic phases of matter based on electronic correlations15, 16. Due to an approximate SU(4) symmetry relating spin and valley, determining which of the components are filled is non-trivial. Past experiments5,6,7,8,9,10,11,12,13, 17 have observed numerous phase transitions between gapped ground states at both integer6,7,8,9, 12, 13, 17 and fractional10, 11, 13 filling. However, these experiments are insufficient to constrain realistic theoretical models, in which the preferred ordering is determined by a combination of the Zeeman energy, which splits the spins; Coulomb interactions and band structure effects, both of which distinguish between the N = 0, 1 orbitals; and several small valley anisotropies which weakly break the valley-SU(2) symmetry18,19,20,21,22,23,24. Indeed, two recent experimental papers explain their data using mutually contradictory single-particle12 and purely interacting pictures10. Constructing a more complete theory of symmetry breaking in bilayer graphene requires experimental determination of the partial filling of each spin, valley, and orbital level, as it evolves with total LL filling.
In this work, we introduce a direct measurement of two out of three of these components, by exploiting the fact that the four valley and orbital components indexed by ξN have different weights on the two layers of the bilayer. We detect this difference in layer polarization capacitively, and use it to infer the fillings ν ξN as a function of both the total electron density and applied perpendicular electric field.
Layer polarization measurements
Our devices consist of hexagonal boron nitride encapsulated B-BLG flakes25 fitted with metal top and bottom gates (Fig. 1a, b). The layer polarization and total charge density are tuned by a combination of applied top and bottom gate voltages (vt and vb), expressed through their symmetric and antisymmetric combinations n0(p0) ≡ ctvt ± cbvb with ct(b) the geometric capacitances between the respective gates and the B-BLG. −n0 and −p0 correspond to the induced charge density and layer polarization in the limit of a perfectly metallic, infinitesimally spaced bilayer. Generically, the physically realized total density (n) and layer density imbalance (p) deviate from this limit, particularly at high magnetic fields. A simple electrostatic model (Supplementary Note 1) shows that these deviations manifest as corrections to the measured gate capacitances (CT and CB) as
where c = (ct + cb)/2 ≈ 1.36 fF/μm2 and c0 is the interlayer capacitance of the bilayer. The antisymmetric combination CA is unique to bilayer systems, vanishing identically in a monolayer system, allowing us to measure the layer polarization p. Measuring CA is technically challenging, as it arises from series combination of the large interlayer capacitance and the comparatively small gate capacitance. It is imperative that the dielectric layers be highly uniform across the device, that the areal mismatch between top and bottom gate be kept , and the geometric capacitance of the two gates be nearly the same; these requirements are difficult to satisfy in conventional two-dimensional electron bilayers but are readily achieved using single-crystal hBN gate dielectrics and atomically thin bilayers.
Figure 1c and e show CS and CA measured at zero magnetic field as a function of n0 and p0. The CS data are dominated by quantum capacitance features of the B-BLG band structure, which features a quadratic band touching at low energies and hyperbolic bands at high energies26. An electric field (p0) opens a band gap with van Hove singularities at the band edges, as can be readily seen in the experimental data (Fig. 1d). Although CS is approximately particle-hole symmetric, significant symmetry breaking contributions are evident. We attribute this to the skew interlayer hopping parameter γ4 in the Slonczewski-Weiss-McClure model for graphite, which breaks particle-hole symmetry by making the lattice non-bipartite (Supplementary Fig. 4)27. CA data, in contrast, reflect the layer-resolved properties of the band wavefunctions (Fig. 1e, f). For p0 ≠ 0, wavefunctions are layer-polarized near the band extrema, so that the first electrons or holes added to the neutral system are added to the corresponding low-energy layer. Reversing p0 inverts the role of the top and bottom layers, inverting the sign of the measured signal with respect to n0. At high overall electron density, the applied p0 is fully screened, so that charge is added symmetrically to the two layers28. The relative layer polarization at different values of n0 can then be extracted by integrating Eq. (2) with respect to n0 (Fig. 1g).
Layer polarization at high magnetic field
Figure 2a shows CS measured in the same device at B = 31 T in the ZLL. We observe insulating states at all integer LL filling factors ν, which are characterized by low capacitance CS and high dissipation (see Supplementary Figs. 1, 2 for dissipation data). Adjusting p0/c at fixed integer ν drives transitions characterized by a spike in CS indicating increased conductivity11 and compressibility, consistent with a closing of the charge gap. Sixteen such phase transitions are evident in the CS data. Similar transitions have been reported in the literature: ref. 6 reported phase transitions at ν = ±2 and p0 = 0, as well as a single phase transition at ν = 0 and finite p0. More recently, the p0 = 0 transitions at ν = ±1, ±3 are evident in ref. 11, while the splitting of the p0 = 0 phase transition at ν = ±2 suggests the formation of a stable, gapped, layer unpolarized state in the region p0 ~0, as was reported in refs. 9, 11. Only in ref. 12 was a potential gapped phase observed at intermediate p0 and ν = 0. However, a unified framework for understanding the diverse competing phases has not yet emerged.
In contrast to layer-insensitive capacitance5, 10, 12 and transport6,7,8,9, 11 measurements, CA provides high experimental contrast throughout the n0 − p0 plane (Fig. 2b). Strikingly, the measured CA falls into discrete levels, corresponding to blue, cyan, orange, and red on the color scale of Fig. 2b. The transitions observed in CS at integer filling fall on lines in the n0−p0 plane along which the sign of CA changes abruptly. To understand the color scale quantitatively, we compute the layer polarization of the ZLL single-particle eigenstates. B-BLG has a 4-site unit cell, with sites A,B in the top layer and A′, B′ in the bottom (see Fig. 2c), and hence the ZLL wavefunctions decompose into their components on the four sublattices, Ψ(x) = (ϕ A (x), ϕ B′(x), ϕ B (x), ϕ A′(x)). The layer polarization is constant across all states in a LL, and independent of spin. It has opposite sign in the two valleys26, so that positive and negative CA correspond to filling valley K and K′, and its magnitude depends on the orbital quantum number, so that α ξNσ = ξα N . At B = 31 T, band structure calculations29 show that α0 = 1 and α1 = 0.63.
As electrons enter LL they contribute a polarization whose magnitude and sign depend on the level being filled. Since , where n0 is very nearly the electron density, we thus interpret red, orange, blue, and cyan as indicating regions where electrons are filling = , , , and type LLs, respectively. This supports a scenario in which, away from phase boundaries, only one of these LLs is filling at each particular (n0, p0). Indeed, numerical calculations (see Supplementary Fig. 5) show that as isospin ξσ fills, around 90% of the electrons enter into either the N = 0 or N = 1 orbital; e.g., either ≳ 0.9 or ≳ 0.9, according to whether the region is red/blue or orange/cyan, respectively.
The polarization of all gapped integer states can now be obtained by summing the level filling sequence starting from the ν = −4 vacuum. Consider ν = 0, where five incompressible states are visible in CS (Fig. 2d). The order in phase I, at large p0 > 0, can be inferred from the observed CA sequence of red, orange, red, orange, implying two N = 0 and two N = 1 states are filled in valley ξ = +. Its layer-inverse occurs in valley ξ = − for large p0 < 0. The phase I states are fully valley polarized, and hence spin and orbitally unpolarized due to Pauli exclusion. Phase II, at p0 near zero, fills levels . A state which fills both orbitals of opposite valleys is consistent with the canted anti-ferromagnetic state8, 22; however, from CA alone we cannot infer whether the spins are polarized or canted. Finally, phase III at intermediate fills levels ±0, ±1, ±0, ∓0.
The orbital and valley filling sequences derived from Fig. 2b provide a more stringent set of constraints on theoretical models than the integer phase transitions alone. For example, a single particle, four-band tight-binding model accounts for the correct number of integer ν transitions12 but fails when compared to the CA filling sequence. The single-particle energies of the ZLL,
are shown in Fig. 3a for B = 31 T. Here, EZ is the Zeeman energy, u is the potential across the bilayer, and Δ10 ∝ γ4B is the splitting between the N = 0, 1 orbitals, which arises because particle-hole symmetry no longer pins the ZLL energies to zero. The interlayer bias u ∝ p0 couples to the differing layer polarization of each state, ξα N , leading to u-tuned crossings at integer fillings; however, the single-particle picture predicts that N = 0 levels fill first for all but the very highest u (Fig. 3b, e), and the transitions would have zero slope in the n0 − p0 plane. This disagrees with the experimental CA data where N = 0 and N = 1 states of the same valley often fill in sequence, and several transition lines have a significant slope.
The departure from the single-particle picture arises from the failure to account for the Coulomb interaction, which both favors sequential filling of N = 0, 1 levels of the same isospin due to exchange18, 19 and penalizes valley polarization due to the interlayer capacitance of the bilayer. However, if Coulomb correlations are made too strong, the first effect dominates and N = 0, 1 orbitals of an isospin always fill sequentially, as shown in Fig. 3d and g. Apparently, competition between the splitting Δ10 and Coulomb correlations is essential.
To determine whether a single model can account for all the observed phase boundaries, we analyze a model that accounts for the single-particle splittings, the SU(4)-invariant screened Coulomb interaction, and several subleading valley anisotropies (Supplementary Note 2). Evaluating the model within a Hartree-Fock approximation allows us to compute the energies of the competing filling sequences, and thereby determine their phase boundaries in the n0 − p0 plane. The model depends on three phenomenological constants (a screening strength, the perpendicular dielectric constant of the BLG, and a valley anisotropy), which we can now estimate by matching the location of the integer transitions and their dependence on an in-plane B-field. The resulting phase diagram is shown in Fig. 3c, and shows good agreement with the CA data of Fig. 2b, including the location of the transitions in absolute units of p0/c and several of their slopes. In our model, each integer state is obtained by fulfilling some number of ξNσ levels; in particular it does not require interlayer coherent phases that spontaneously break the valley U(1) × U(1) symmetry, in contrast to theories that predict such phases at ν = ±1, p0 = 0. We also predict that the ν = 0 phase III observed at ν = 0 likely hosts helical edge states similar to those recently described in twisted bilayer graphene30. This state is stabilized by the single-particle anisotropy Δ10 and antagonized by the Coulomb interactions, suggesting it could be further stabilized in devices with stronger screening due to proximal metal gates.
Despite good overall agreement, there is an interesting qualitative discrepancy between the Hartree-Fock analysis and the data. In the experiment, the slope of the phase transition line between −2 < ν < −1 (see Fig. 4a, boundary (ii)–(iv)) is significantly larger than the slope of the adjoining transition across −3 < ν < −2 (boundary (i)–(iii) of Fig. 4a). Within the Hartree-Fock approximation, the slopes are identical, and in fact any model which neglects scattering (e.g., “Landau-level mixing”) between the N = 0, 1 orbitals has a particle-hole symmetry ν + 2 → −(ν + 2), u − u*(ν = −2) → −(u − u*(−2)), forcing the two boundaries to mirror each other. To account for the asymmetry, we instead find the ground state of the model’s Hamiltonian using the multicomponent infinite-density matrix renormalization group, which takes full account of correlations31. Allowing LL mixing produces a kink in the slope at ν = −2, as experimentally observed (Fig. 4b). LL mixing is known to generate effective three-body interactions, which stabilize fractional non-Abelian phases32, 33; our results suggest that these interactions may be stronger in BLG than in conventional semiconductor quantum wells.
In conclusion, we have described a new experimental technique to determine the layer polarization of van der Waals bilayers and used it to constrain a detailed model of symmetry breaking in the bilayer graphene ZLL. Our technique is readily applicable to quantitatively probe layer, valley, and spin polarization in other atomic layered materials, including twisted bilayer graphene and both homobilayer and heterobilayer of transition metal dichalcogenides.
Bilayer graphene samples encapsulated in hexagonal boron nitride were fabricated using a dry transfer method25. Particular care is taken to ensure that the top and bottom hBN flakes are of the same thickness, measured by atomic force microscopy to be 19 and 20 nm, respectively. During fabrication, care is also taken to minimize the area of graphene bilayer gated by only one of the two gates, as single-gated areas contribute a systematic error to the measured CA signal proportional to the area and to CS. Anticipating ≲ , we ensure that the areal mismatch between bottom-gated and top-gated areas is less than .5%.
Capacitance measurement electrical schematic
Capacitance measurements were made using a cryogenic impedance transformer based on an FHX35X high electron mobility transistor34 in a bridge configuration connected to the bilayer graphene ohmic contacts (see Fig. 5). vg sets the transistor bias point and vd adjusted to be sufficiently low that no sample heating is observed. To measure CS(A), two synchronized and nearly equal-magnitude AC signals (δVEX) are applied to the top and bottom gates, whose relative magnitude is chosen to match the ratio of geometric capacitances ct/cb extracted from the DC characteristics of the device. The signals are applied in phase for CS and out of phase for CA. A third AC signal is applied to a standard capacitor Cstd with amplitude and phase that null the signal at the input of the cryogenic amplifier, and the capacitance and dissipation determined from the relation of the AC signals. Cstd was measured to be 404 fF during the cooldown in which the data of Fig. 1 were measured. We used this value to determine CS and CA shown in subsequent figures, although in our experience Cstd can vary by up to 20 fF from cooldown to cooldown, thus introducing a systematic uncertainty of approximately 5% in the capacitance shown in Figs. 2–4. All data shown are acquired off-balance, by monitoring the voltage at the balance point as DC values of the gate voltages are changed. Data in Fig. 2a and b are measured at 67.778 kHz using a 10 mV variation of n0/c and of p0/c, respectively.
Interpretation of CA as a thermodynamic derivative requires that the sample is sufficiently conductive to fully charge over a time scale comparable to the inverse measurement frequency35. At low temperature and high magnetic fields, our sample becomes strongly insulating at integer filling factors, precluding this condition being satisfied. Failure to charge manifests as an increase in the out of phase, dissipative signal in the capacitance, allowing us to monitor charging across the parameter range. In Figs. 2b and 4a, regions in which the sample does not charge are masked in black, and dissipation data for all data sets is shown in Supplementary Figs. 1 and 2.
The data that support the findings of this study are available from the corresponding author upon reasonable request.
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We acknowledge helpful discussions with E. Berg, L. Levitov, A. Macdonald, and I. Sodemann. The work at UC Santa Barbara was funded by NSF under EAGER DMR-1636607. The work at MIT was funded by the BES Program of the Office of Science of the US DOE, contract no. FG02-08ER46514, and the Gordon and Betty Moore Foundation, through grant GBMF2931. Work at Columbia was funded by the NSF MRSEC under DMR-1420634, and by the US ONR under N00014-13-1-0662. K.W. and T.T. acknowledge support from the Elemental Strategy Initiative conducted by the MEXT, Japan, and JSPS KAKENHI grant number JP15K21722.
Electronic supplementary material
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.