Introduction

Due to its lattice structure and position of the Fermi energy, the low-energy electronic excitations of graphene are described by an effective field theory that is Lorentz invariant1. Unlike Galilean invariant theories such as Fermi Liquids2 whose main relevant parameter is the effective mass, Lorentz invariant theories are characterized by an effective velocity. Because of this, an increase of electron-electron interactions induces an increase of the Fermi velocity, vF, in contrast to Fermi liquids, where the opposite trend is true3. In the case of graphene, when electron-electron interactions are weak4, vF is expected to be as low as 0.85×106 m/s, whereas, for the case of strong interactions5, vF is expected to be as high as 1.73×106 m/s.

Recently, Fermi velocities as high as ~3×106 m/s6 have been achieved in suspended graphene through a change of the carrier concentration n6,7,8,9. However, because this dependence is logarithmic, n needs to be changed by two orders of magnitude in order to change the velocity by a factor of 3. This implies that it is unpractical to use n as a way to engineer vF, let alone the fact that one should first realize suspended graphene in the device6. Several other routes have also been proposed to engineer vF in graphene via the electron-electron interaction, including modifications of: a) curvature of the graphene sheet10; b) periodic potentials11; c) dielectric screening12,13,14. While the former two also substantially modify the starting material, the latter simply modifies the effective dielectric constant, ε, making it more appealing for device applications15. Despite this advantage, no systematic study of how to engineer vF by changing ε exists to date. Here we provide a new venue to control the Fermi velocity of graphene using dielectrics, while keeping n constant.

Results

We perform such a study using three single-layer graphene samples, which were prepared by chemical vapor deposition (CVD) on Cu, followed by an in situ dewetting of Cu on quartz (single crystal SiO2)16 or a transfer onto hexagonal boron nitride (BN)17 and by epitaxial growth on 4H-SiC(000-1)18. Figures 1A and 1B show angle-resolved photoemission spectroscopy (ARPES) intensity maps measured near the Brillouin zone corner K along the Γ-K direction for the two CVD grown samples, which constitute the first report on Dirac quasiparticle mapping from these samples. Following the maximum intensity, one can clearly observe almost linear energy spectra, characteristic of Dirac electrons19. The momentum distribution curves (MDC), intensity spectra taken at constant energy as a function of momentum, are shown in Fig. 1C. In addition to being proportional to the imaginary part of the electron self-energy, the MDC spectral width provides information on the sample quality. A clear increase of the width is observed by changing the substrate from SiC(000-1) via BN to quartz, a trend that is in overall agreement with the theoretical expectation that the electron self-energy should vary with the inverse square of the dielectric screening20, as later discussed. The quartz sample here used constitutes a substantial improvement over a previous experiment on a similar substrate21 (compare 0.19 Å−1 (red line) versus ~0.7 Å−1 (gray-dashed line)). The much improved data quality allows for a detailed self-energy analysis and consequent extraction of important parameters such as vF.

Figure 1
figure 1

ARPES intensity maps of graphene on quartz and BN.

(A–B) Normalized and raw ARPES intensity maps of graphene/quartz (panel (A)) and graphene/BN (panel (B)), respectively. The red and dark-yellow lines are the dispersions, obtained by fitting momentum distribution curves (MDCs). (C) MDCs at EF for graphene on SiC(000-1) (blue line), BN (dark-yellow line), quartz (red line) and SiO221 (gray-dashed line).

To understand how the dielectric substrate affects the electronic properties, in Fig. 2, we show the energy vs. momentum dispersions for graphene on three different substrates, SiC(000-1), BN and quartz, obtained by fitting the MDC spectra. The observed dispersions exhibit two distinctive features. First, the measured dispersions deviate from linearity with an increased slope around ~0.5 eV for all the samples (compare experimental data to dashed gray lines in Fig. 2A). As the substrate is changed from SiC(000-1) via BN to quartz, corresponding to a decrease of the dielectric screening, the departure from linearity at high energy becomes more pronounced. Second, the direct comparison between experimental dispersions and ab initio calculations for the two extreme cases ε = 15 (suspended graphene) and ε = ∞4 shows another substrate-dependence (Fig. 2B). Upon changing the substrate, the slope increases approaching the dispersion for ε = 1. The deviation from linearity and the enhancement of the slope result in a reshape of the typical conical dispersion, in a similar fashion as reported for other charge-neutral graphene samples6,12 (see cartoons in the inset of Fig. 2A: from left to right). We note that the largest upturn for graphene/quartz cannot be explained by: a) resolution, which typically results in the deflection of MDC peaks near EF to lower momentum and would involve a much smaller effect by an order of magnitude (≤a few tens meV)22; b) the presence of other bands with a different azimuthal orientation, which would cause instead an abrupt increase and a significant asymmetry of the MDC width at the upturn energy.

Figure 2
figure 2

Experimental and theoretical energy spectra for different dielectric constants.

(A) Experimental dispersions for graphene on SiC(000-1) (blue line), BN (dark-yellow line) and quartz (red line). The gray-dashed lines are guides to the eyes. The insets are cartoons for the electron band structure of graphene with weak (left) and strong (right) electron-electron interactions. The data are shifted along the x-axis. (B) The direct comparison of experimental dispersions with theories: ε = ∞ (magenta line)4 and ε = 1 (cyan line)12.

Discussion

To quantify the effect of dielectric substrates on the electron-electron interactions and vF, we adopt the standard self-energy analysis to extract self-consistently the strength of the electron-electron interactions and ε1,12,23,24. Figure 3A shows the difference between measured dispersions, E(k) (from Fig. 2A) and the theoretical dispersion for ε = ∞, ELDA(k) (shown in Fig. 2B). Assuming that electron-electron interactions are effectively screened for ε = ∞, the E-ELDA curve can be considered a good measurement of the difference between the self-energy and its value at EF. To fit these curves, we use the marginal Fermi liquid self-energy function as previously reported12,23 with an analytic form of ln(kC/(kkF)) (dotted lines in Fig. 3A). Here, α is a dimensionless fine-structure constant (or the strength of electron-electron interactions) defined as 23, v0 the Fermi velocity for ε = ∞, 0.85×106 m/s4, kC the momentum cut-off, 1.7 Å−1 and kF the Fermi wave number. An overall good agreement with the experimental data is observed allowing us to extract important parameters such as ε and α for graphene on each substrate. For graphene on SiC(000-1) and BN, we obtain ε = 7.26±0.02 (α = 0.35) and ε = 4.22±0.01 (α = 0.61), respectively. The extracted value for graphene on BN is in agreement with the standard approximation ε = (εvacuumsubstrate)/2 = 4.02 and 3.05, where εvacuum = 1 and εsubstrate = 7.04 (for out-of-plane polarization) and 5.09 (for in-plane polarization) in the low frequency limit (static dielectric constant) for hexagonal-BN25. Similarly, the obtained value for graphene on SiC(000-1) is close to a previous report12. The apparent discrepancy with the latter (compare ε = 7.26±0.02 in this work with 6.4±0.1 in reference12) is due to the different choice of reference band (or so-called bare band). Specifically, in this work, ELDA is used as the bare band, whereas, in reference 12, the bare band is approximated by a straight line. Finally, for graphene/quartz, we obtain ε = 1.80±0.02 (α = 1.43), which is smaller than the expected value of ε = 2.4526, instead closer to the experimentally extracted value for suspended graphene (~2.2)6. This observation, together with the similar energy-momentum dispersion relation at high binding energy to the theoretical one for suspended graphene (Fig. 2B), points to a very weak effect of the substrate. This is likely a consequence of the different sample preparation method adopted here (see Methods section).

Figure 3
figure 3

Fermi velocity and the strength of electron-electron interactions.

(A) E-ELDA dispersions for graphene on SiC(000-1) (blue line), BN (dark-yellow line) and quartz (red line). (B) Fermi velocities as a function of ε. The dashed line is a theoretical curve for vF, which is inversely proportional to ε6,23. Filled symbols correspond to experimental results, while empty symbols to theoretical values. ε = 2.45 for G/SiO226 is obtained from the standard approximation, ε = (εvacuumsubstrate)/2 (see text). (C) The ratio of vF, the renormalized Fermi velocity due to electron-electron interactions, to v0 = 0.85×106 m/s, the bare Fermi velocity in the LDA limit where ε = ∞4, as a function of α. The dashed line is the fit given by vF/v0 = 1-3.28 α{1+(1/4) ln[(1+4 α)/4 α]-1.45}31 for charge neutral graphene. The inset is the relation between α and ε, where the dashed line is 23.

In Fig. 3B, we show the measured vF as a function of the extracted ε (see also Table 1). Results from a suspended sample6 and another graphene/SiO2 sample21 are also plotted for comparison. Upon decreasing ε from ∞ to 7.26 and 4.22, vF is enhanced from its LDA limit of 0.85×106 m/s (cyan triangle in Fig. 3B) to 1.15±0.02×106 m/s (blue circle in Fig. 3B) and 1.49±0.08×106 m/s (dark-yellow circle in Fig. 3B), by 35% and 75%, respectively. Surprisingly, when ε is further decreased to 1.80, a dramatic enhancement of vF up to 2.49±0.30×106 m/s (red circle in Fig. 3B) is observed. Such enhancement corresponds to a 190% increase from its bare value and represents the highest value reported for graphene on any substrate27,28,29. Interestingly, this velocity is comparable to the value measured for suspended graphene (green square in Fig. 3B)6. Clearly, a 1/ε dependence of vF is observed (dashed line in Fig. 3B) in agreement with the theoretical prediction6,23. Our result constitutes the first observation of a power law dependence of the Fermi velocity on the dielectric constant at fixed n. This power law dependence allows one to achieve, by a smart choice of dielectric, a high value of vF that cannot be attained otherwise by changing n6.

Table 1 Fermi velocity (vF), dielectric constant (ε) and fine structure constant (α) of graphene on each substrate

We note that CVD graphene on quartz (red circle in Fig. 3B) exhibits higher vF than exfoliated graphene on amorphous SiO2 (gray square in Fig. 3B) with the same stoichiometry as quartz. This is a consequence of different sample preparation process and is due to the larger presence of impurities in the exfoliated sample, as suggested by the extremely broad spectra (see gray dashed line in Fig. 1C). Therefore, although, in theory, one should expect smaller vF due to screened electron-electron interactions from impurity13, one should be cautious in extracting meaningful parameters from these data. We also note that ab initio GW calculations5 (magenta triangle in Fig. 3B) underestimate vF of suspended graphene. This may be due to the finite k-point sampling inherent in such calculations, or it could also be an indication of the need to add higher-order terms in the self-energy calculation by the GW-approximation.

In Fig. 3C, we plot the ratio between vF and v0, the expected Fermi velocity in the fully screened case (ε = ∞), as a function of α. As the strength of electron-electron interactions is increased, vF is also enhanced. This is in striking difference with the standard Fermi liquid picture, where vF is expected to decrease with increasing α30. On the other hand, the observed behavior is consistent with previous theoretical studies for graphene in the case of specific electron-electron interactions30,31 (dashed line in Fig. 3C) exhibiting the characteristic self-energy spectrum analogous to a marginal Fermi liquid1. As a result, the departure from the Fermi liquid picture becomes more important with increasing electron-electron interactions or decreasing dielectric screening (see the relation between α and ε in the inset of Fig. 3C). Additionally, the observation of α value close to 1 (neither α1 nor α1) for graphene/quartz may indicate that a full theoretical treatment beyond the random-phase approximation1 may be required to understand this sample and/or suspended graphene6.

The very good agreement with theoretical predictions23,31 for both vF versus ε (Fig. 3B) and vF versus α (Fig. 3C) confirms that the dielectric constants obtained by the self-energy analysis are self-consistent. Finally the experimentally determined ε can largely account for the relatively broad MDCs observed for graphene on quartz (Fig. 1C), as compared to graphene on BN and SiC(000-1). For ε values of 1.80, 4.22 and 7.26, for graphene on quartz, BN and SiC(000-1) respectively, the MDC widths, expected to vary with the inverse square of the dielectric screening20, should be roughly 16 and 5 times broader for graphene on quartz and BN than graphene on SiC(000-1), in line with the experimental observation (see, for example, Fig. 1C). We stress that, contrary to a Fermi liquid system, the broader MDC spectra observed for graphene/quartz do not necessarily imply decreased transport properties. On the contrary, the enhanced α, the primary cause of the broad spectra, give rise to an enhancement of Fermi velocity, which is ultimately one of the most important parameters for device applications.

In conclusion, we have unveiled the crucial role of dielectric screening in graphene to control both Fermi velocity and electron-electron interactions. Additionally, we have shown that graphene, in its charge neutral state, departs from a standard Fermi liquid not only in its logarithmic energy spectrum as previously discussed12, but also in the way that vF is modulated by the strength of electron-electron interactions. This dependence provides an alternative way to engineer Fermi velocity for graphene on a substrate by modifying the dielectric substrate. This approach can also be applied to charge-doped graphene and other two-dimensional electron systems such as topological insulators32 that can be grown or transferred to dielectric substrates.

Methods

Graphene samples were prepared in three different ways: epitaxial growth on the surface of a 4H-SiC(000-1) substrate; chemical vapor deposition (CVD) growth on a Cu film followed by a transfer onto the surface of boron nitride17; and CVD growth followed by in situ dewetting of Cu layer in between graphene and a single crystal SiO2 (namely quartz which is different from amorphous SiO2 on an Si substrate, the widely used substrate for exfoliated graphene27) substrate16. The later procedure is clearly different from the standard method of exfoliating graphite followed by deposition onto the amorphous SiO2 layer21. This results in a reduced effect of the substrate that is suggested by the enhanced height variation with respect to the substrate compared to the sample prepared by the exfoliation and deposition16,33. The resulting graphene is more decoupled from the substrate as supported by several features such as Fermi velocity, dielectric constant and the electron band at higher energies closer to suspended sample.

In order to remove any residue including Cu and PMMA, a precursor to grow CVD graphene and a polymer to transfer graphene, respectively, we heated the sample to 1000 oC in ultra-high vacuum. The removal of Cu is confirmed by: (a) optical microscopy showing a cleaner image without residual Cu once the sample has been heated; (b) absence of related Cu features in the ARPES spectra such as 3d electrons at 3.0 eV and 3.5 eV below Fermi energy and 4s free-electron-like state with a band minimum at 0.25 eV below Fermi energy34.

High-resolution ARPES experiments have been performed at beamline 10.0.1.1 of the Advanced Light Source at Lawrence Berkeley National Laboratory using 50 eV photons at 15 K. Energy and angular (momentum) resolutions were set to be 22 meV and 0.2 o (~0.01 Å−1), respectively.