Wrinkles Versus Domes: An Experimental Test of Curvature-Bound States in Graphene

Electrons in graphene behave like massless relativistic particles, and they pass straight through ordinary electric barriers (Klein tunneling). This is why electrostatic quantum dots work poorly in graphene. A body of simulation work proposes a different trap: bend the sheet. On a curved surface, an electron's internal two-component state must keep rotating to stay aligned with the local surface. The correction that does this, called the spin connection, enters the equations of motion much as a magnetic field would. Simulations predict that a curved bump traps a small fraction of an incoming electron wave, about 0.1%. They also predict that this trapped fraction grows as a power of the average curvature, with exponent about 0.81 on a graphene lattice, and that the bump shows discrete energy levels in zero magnetic field (Confining massless Dirac particles in two-dimensional curved space). None of this has been measured.

The obvious experiment is to make bumps, measure their local spectra with a scanning tunneling microscope, and look for discrete peaks. That experiment cannot settle the question. Bending graphene also stretches its bonds, which produces a strain-induced "pseudomagnetic" field (Interplay between real and pseudomagnetic field in graphene with strain). It also misaligns the carbon orbitals, which produces a scalar potential that pools charge (Graphene as an electronic membrane). Either effect alone can produce discrete peaks. We propose an experiment designed to tell the three mechanisms apart, built on one geometric fact: two shapes can have the same bending yet differ in intrinsic curvature.

Core Claim

For a two-dimensional surface, the Ricci scalar that controls the predicted trapping law equals twice the Gaussian curvature, $R = 2K$. Here $K$ is the product of the two principal curvatures at a point. A one-directional wrinkle is bent but intrinsically flat ($K = 0$), like a rolled sheet of paper. A dome has $K > 0$. The spin-connection mechanism depends only on the intrinsic geometry, so it predicts no bound states on a smooth wrinkle and bound states on a dome.

The two competing mechanisms do not vanish on wrinkles. The scalar potential scales with the square of the mean curvature, $\Phi \propto -(\nabla^2 h)^2$, where $h(x,y)$ is the height map (Graphene as an electronic membrane). This is nonzero on a wrinkle. The pseudomagnetic field follows the in-plane strain, and the strain depends on how the sheet is clamped as well as on its shape.

Our claim is that a matched set of wrinkles and domes with equal mean curvature, measured at zero field, separates the three mechanisms. Varying the clamping of identical domes adds a second, independent separation. Together these measurements give the first experimental test of the $P_c \propto \langle R \rangle^{\alpha}$ law, where $P_c$ is the trapped fraction and $\langle R \rangle$ the average Ricci scalar.

There is supporting simulation evidence for this asymmetry. In a magnetic field, curvature shifts Landau levels with a coefficient of only $\xi_0 \approx -0.10$ for 1D wrinkles but $-0.78$ for biaxial ripples (Landau levels in wrinkled and rippled graphene sheets). Geometry with $K \neq 0$ therefore matters far more than bending alone.

Research Program

Making the simulated quantity measurable. The predicted law concerns the trapped fraction of a wavepacket, which no microscope measures. In the curved-space quantum lattice Boltzmann solver, a grid-based method for propagating the Dirac equation on curved surfaces (Confining massless Dirac particles in two-dimensional curved space), we will compute two quantities for each shape: the trapped fraction, and the spectral weight of the discrete peaks in the local density of states summed over the bump. The local density of states is what tunneling spectroscopy measures. If the two track each other across shapes, the measurable weight inherits the predicted exponent. If they do not, we will test the law on the weight and state that the link is unestablished.

A switchable model. We will build one model of a measured surface in which each mechanism can be turned on or off: the metric plus spin connection, the strain pseudomagnetic field, the curvature-dependent Fermi velocity (Tunable Dirac points and zero-energy modes in periodic curved graphene superlattices), and the scalar potential $\Phi$. Its input is the measured height map, not an idealized Gaussian or sinusoid. The in-plane strain comes from an elastic relaxation of that map under the actual boundary condition (pillar-supported or bubble-enclosed). We will check the full model against an atomistic tight-binding calculation on the same relaxed lattice, which includes all mechanisms without separating them.

The curved-space solver becomes unstable above roughly 15% deformation (Landau levels in wrinkled and rippled graphene sheets). We will therefore keep the ratio of bump height to width below about 0.15 and use tight-binding results beyond that.

Samples and measurement. We will place monolayer graphene on substrates patterned with two families of features: elongated ridges, which produce wrinkles, and round pillars or trapped bubbles, which produce domes. Lateral sizes will range from about 8 to 20 nm, with matched mean curvature across families.

Lateral size is set by the energy resolution. The level spacing in a region of size $L$ is roughly $\hbar v_F / L$, about $0.66$ eV·nm divided by $L$, or 33 to 80 meV over this range. These spacings must exceed the 25–38 meV broadening that was needed to fit tunneling spectra of a flat monolayer (Acquisition and Analysis of Scanning Tunneling Spectroscopy Data − WSe₂ Monolayer). They must also exceed the ±30 meV charge-puddle fluctuations expected on SiO₂ (Graphene as an electronic membrane).

Spectra will be taken at 4 K and zero field. We will use the variable tip-height ramp protocol of the WSe₂ study to capture weak in-gap features over several decades of current. Topography and spectroscopy will be recorded on the same grid.

Calibration. Each mechanism variant gets the same small set of free nuisance parameters: instrumental broadening, a tip-induced energy offset, and an overall Fermi-velocity scale. These will be fitted with a tree-structured Parzen estimator under an equal evaluation budget, following the simultaneous global calibration used for a 41-parameter polymer model (Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization: A Case Study on Pebax Polymer). Calibration uses a training subset of features. All comparisons are reported on held-out features.

Baselines

Naive reference. The do-nothing reference is a flat-graphene Dirac model with no curvature terms. Its only input is the measured local doping, taken from the shift of the Dirac point in nearby flat regions. This model should predict featureless, roughly linear spectra on every bump. Any mechanism model must reduce the held-out spectral error well below this level.

Strongest physical competitor. The second baseline is the electronic-membrane model, which adds the scalar potential $\Phi \propto -(\nabla^2 h)^2$ computed from the measured height map (Graphene as an electronic membrane). It is the strongest competitor because it has already matched real data: it gives a simulated height spread of 0.22 nm against 0.19 nm measured on SiO₂, and accounts for the observed charge puddles. It is also the only competitor that predicts levels on wrinkles. It is calibrated with the same parameters and budget as every other variant.

Strain-only model. The third baseline treats the sheet as flat and confines electrons only by the strain-induced pseudomagnetic field, driven by the relaxed strain from our measured map (Interplay between real and pseudomagnetic field in graphene with strain). That paper reports fields above 10 T and bound dot states, but no measurement. We found no reported spectroscopic result on Lacuna for either strain-only or curvature-trapping models on real graphene. The comparison numbers will therefore be generated here, by these baselines run under identical conditions.

Ablation of the key ingredient. The ablation isolating our hypothesis is the full model with only the spin connection and metric removed. Everything else stays: the same input surfaces, relaxation, nuisance parameters, and calibration budget. Beating the baselines means the full model achieves lower held-out spectral error and lower peak-energy error, in meV, than both the scalar-potential and strain-only models on domes. It must achieve no worse error on wrinkles. Its advantage over the ablation must be concentrated on domes and absent on wrinkles.

Why not just…?

Why not just look for discrete peaks on bumps and count them? Discrete peaks are expected under all three mechanisms, and puddle-induced levels of tens of meV are expected even on random ripples (Graphene as an electronic membrane). A positive peak detection would therefore not distinguish curvature trapping from charge pooling. The wrinkle-versus-dome contrast is what separates them.

Why not just run atomistic tight-binding on the measured topography? This is a real threat, since tight-binding contains all the physics automatically and may fit the data best. We use it as the reference that validates our switchable model. Its weakness is that it cannot say which term produced a level. If the switchable model matches tight-binding and the data, the switches supply the attribution. If it does not, that failure is itself a result: the curved-space continuum picture would be missing something tight-binding captures.

Why not just train a neural classifier on simulated spectra, as was done for nematic order? A one-hidden-layer network trained on 6,000 simulated 16×16 maps reached over 95% accuracy on synthetic data and over 90% confidence on real data (Detecting nematic order in STM/STS data with artificial intelligence). Gaussian mixture clustering also separated two nematic types in twisted bilayer graphene (Unsupervised learning of two-component nematicity from STM data on magic angle bilayer graphene). Such a classifier inherits the assumptions of its simulator. With only tens of bumps per sample, the small-sample problem noted in the bilayer study applies. We will train such a classifier as a secondary check, but use the physically motivated geometric contrast as the primary test.

Why not just confine electrons electrostatically or with planar strain, and skip curvature? Electrostatic barriers fail in graphene because of Klein tunneling. Planar strain is close to unrealizable as truly planar in a membrane with bending rigidity near 1 eV. In any case, the question here is whether curvature itself traps electrons, so these routes do not answer it.

Proposed Experiments and Decisive Checks

The first decisive check is the wrinkle–dome test. Domes and wrinkles are matched in mean curvature. If the spin-connection mechanism matters, held-out spectra should show discrete levels with larger spectral weight on domes, and the ablation should fail mainly on domes. If wrinkles show levels as strong as domes and the scalar-potential baseline fits both, the evidence favors pooled charge rather than intrinsic curvature as the trap.

The second check varies clamping at fixed shape. Pillar-supported and bubble-enclosed domes with matched height profiles have different strain fields. A large spectral change between them implicates the pseudomagnetic field.

The third check uses symmetry and atomic-scale signatures. The scalar potential shifts levels asymmetrically about the Dirac point. Strain fields produce levels with strong sublattice polarization, visible in atomically resolved maps. Spin-connection levels should be roughly symmetric about the Dirac point and not sublattice polarized.

The final test is the curvature scaling law. Across domes of varied height-to-width ratio, we will fit the summed bound-state weight against $\langle R \rangle$ from the measured topography. We then ask whether the exponent is consistent with the predicted 0.81, after confirming in simulation that this weight tracks the trapped fraction. A clear miss with good spectral fits would mean the idealized trapping law does not survive real topography and disorder. That is also worth knowing.