Twisted bilayer graphene is two sheets of carbon rotated slightly against each other. Its electrons are extremely sensitive to the rotation (twist) angle. Real devices never have one angle everywhere: they break into patches of slightly different twist separated by boundaries called domain walls. A clean theory treats each wall as a quantum "step" that electrons must climb. The step exists because the energy bands on the two sides sit at different energies. The theory predicts two electrical fingerprints of twist disorder: a gap in conductance at zero voltage, and peaks in current noise Transport across twist angle domains in moiré graphene. In that picture the step height is simply the band offset between two ideal, uniform, non-interacting domains.
That assumption skips a basic fact about connected metals. When two regions with different band positions touch, charge flows between them until a single electrochemical potential is shared. The resulting electrostatic dipole partly cancels the bare offset, and in flat-band systems, where the density of states is huge, this cancellation could be nearly complete. This proposal asks one question: once electrons are allowed to redistribute self-consistently, and the wall is given its real relaxed atomic structure and a realistic treatment of electron exchange, how large is the step, and do the predicted transport gap and noise peaks survive? We answer it with density functional theory (DFT), a first-principles method that computes the electron density and effective potential of a material from its atomic positions. We use it as the reference for a transport calculation across explicit, atomistically relaxed domain walls.
Core Claim
The step-potential fingerprints of twist disorder are governed by the screened offset, not the bare band offset. The screened offset depends on each domain's density of states near the Fermi level and on the width of the wall. It therefore cannot be read off from calculations on uniform crystals. Concretely, we hypothesize the following. Near the magic angle (about 1.08°), self-consistent charge transfer shrinks the effective step well below the bare offset. The zero-bias gap should then scale with twist mismatch much more weakly than linearly. The noise peaks tied to the band-edge singularities should stay pinned to each domain's own singularity energies, but be broadened by the finite wall width. A falsifiable version of the claim is that a model built by stitching together the bands of two uniform-angle DFT calculations will mispredict the zero-bias gap by more than the gap's thermal resolution at a few kelvin. If it does not, the step-potential picture is quantitatively vindicated at first-principles level. That would also be a useful result, and we would report it as such.
Why This Is Not Already Done
The step-potential paper is purely a continuum model. It explicitly flags sharp walls and semi-infinite uniform domains as idealizations Transport across twist angle domains in moiré graphene.
Several first-principles methods reach moiré scale, but they have only been applied to uniform, periodic twist. The linear-scaling OpenMX method handles the 11,164-atom magic-angle cell and captures corrugation and intervalley gaps An $O(N)$ $Ab~initio$ Calculation Scheme for Large-Scale Moiré Structures. A deep network that predicts hybrid-functional Hamiltonians shows that exact exchange widens the flat bands roughly tenfold A deep equivariant neural network approach for efficient hybrid density functional calculations. DeepH-2 reaches sub-meV Hamiltonian accuracy on graphene DeepH-2: Enhancing deep-learning electronic structure via an equivariant local-coordinate transformer. None of these computes open-boundary transport, and none studies a wall between two angles.
The obvious combination is "run DeepH on a big wall, then compute transmission." It has a specific flaw that motivates this proposal. DeepH-type models rely on electronic nearsightedness, meaning each Hamiltonian block depends only on nearby atoms. The long-range electrostatic dipole that forms when two domains exchange charge is exactly the nonlocal effect such models cannot learn from periodic training cells. The new ingredient here is to separate the problem into two parts. One is the local Hamiltonian, learned and validated against DFT. The other is an explicit self-consistent electrostatic correction solved on the whole device, with DFT-level SCF calculations on smaller walls as the check.
Research Program
The work has four stages.
Geometry. We build domain-wall structures in which the twist changes from θ₁ to θ₂ across a finite region. The structure is periodic along the wall so that a transport calculation is well defined. The two moiré lattices are not commensurate along the wall direction. We restore periodicity with a small uniaxial strain, and we check that the strain is harmless by comparing two different commensurate approximants for the same angle pair. The structures are relaxed with DRIP, a DFT-fitted interlayer potential with a dihedral-angle correction that fixes the out-of-plane force errors of Kolmogorov–Crespi potentials Dihedral-angle-corrected registry-dependent interlayer potential for multilayer graphene structures. We cross-check the relaxed geometry with the hNN-Gr$_x$ neural potential Hybrid neural network potential for multilayer graphene. Both potentials are on OpenKIM. They supply the atomic structure only; they give no electronic information.
Hamiltonian. We fine-tune DeepH-2 on OpenMX calculations of relaxed twisted bilayers, adding strained and wall-containing cells at large angles (roughly 2–4°, a few thousand atoms each). At these angles direct OpenMX SCF runs remain affordable, at 14.6 core-hours per SCF loop even at the 11,164-atom scale. A hybrid-functional variant uses the DeepH-hybrid approach. We report the validation error of both models on held-out wall configurations, not only on uniform crystals, because generalization to walls is exactly what has not been shown.
Electrostatics. We run a self-consistent loop on the full device. The learned Hamiltonian gives a local density of states. A Poisson solve that includes the hexagonal boron nitride dielectric environment and a gate gives the Hartree potential. That potential is added to the on-site terms, and the loop repeats until the Fermi level is uniform. The critical check is to run this reduced scheme and full OpenMX SCF on the same large-angle wall cells and compare the resulting potential profiles across the wall.
Transport. We compute transmission $T(E, k_\parallel)$, where $E$ is energy and $k_\parallel$ is momentum along the wall, using recursive Green's functions with semi-infinite leads of uniform θ₁ and θ₂. Conductance follows from the Landauer formula and the Fano factor from $\sum T(1-T)/\sum T$. A secondary branch applies the same pipeline to twisted double bilayer graphene. There, wall polarization measured by piezoresponse microscopy Co-Dominant Piezoelectric and Flexoelectric Effects in Twisted Double Bilayer Graphene implies bound charge that would add its own term to the step.
Baselines
Continuum step model (naive reference). This is the continuum two-band step model of Transport across twist angle domains in moiré graphene, with its sharp wall and bare band offsets, re-implemented with the same angle pairs and the same energy grid. It is the claim being tested. The file gives no numerical transmission or gap values for it: no reported result found on Lacuna for absolute transmission or gap magnitude in that paper. We therefore regenerate its curves ourselves.
Locally-uniform DFT model (strongest competitor). This is the cheap alternative a skeptical reviewer would propose, and it is the one that could genuinely compete. Each domain's bands come from a periodic calculation at its own angle: OpenMX or DeepH-2 for the semi-local functional, DeepH-hybrid for the hybrid functional. The two band structures are then stitched into a step and solved with the same transfer-matrix code. Hamiltonian accuracies are MAE 0.12 meV for DeepH-2 on graphene and about 0.2 meV for DeepH-hybrid, so band input errors are far below the effects of interest. We give this baseline the same functional and the same relaxed uniform-domain geometries as the full method. The only thing it lacks is the explicit wall and the self-consistent electrostatics. Our method beats it if the full calculation's zero-bias gap, or its noise-peak energies, differ from it by more than about 0.5 meV (roughly $k_BT$ at a few kelvin), and if the difference grows with twist mismatch in a way screening can explain.
Relaxed tight-binding transport. This baseline uses a standard Slater–Koster graphene tight-binding model on the same DRIP-relaxed walls, with the same Green's-function code. It isolates what the first-principles Hamiltonian adds beyond geometry.
Ablations. Three ablations isolate the key ingredient. We switch off the Hartree loop (DeepH only), use rigid instead of relaxed geometry, and swap the semi-local functional for the hybrid one.
Why not just…?
Why not just use the continuum model with better parameters? Continuum models omit the intervalley coupling and corrugation effects that the OpenMX study shows reshaping the flat bands (bandwidth 21.0 meV rigid versus about 13.2 meV corrugated). Their parameters are also tuned to calculations whose bandwidth the hybrid study says may be off by an order of magnitude (about 4.1 versus 41.1 meV). These figures come from different setups, and that disagreement is itself a reason to compute functional dependence directly. Most importantly, no choice of parameters adds charge screening at the wall. The "no Hartree" ablation measures exactly how much that omission costs.
Why not just stitch uniform-angle DFT bands into the step model? This is the real threat. If screening turns out to be weak, because the hBN and gate electrostatics suppress charge transfer or because walls are wide, this cheap model will match the full calculation. The proposal would then reduce to validating it. The second baseline comparison settles the question directly, and we will report the answer either way. A positive outcome would give experimentalists a license to use the cheap model.
Why not just use relaxed tight-binding with transport, as is routine for moiré devices? Tight-binding can capture geometry. But its on-site energies do not respond to charge redistribution unless a Hartree term is added by hand, and its hoppings do not know about exact exchange. We will include it as a baseline. If it matches DeepH with the electrostatic loop once the same Hartree loop is bolted on, we will conclude the first-principles Hamiltonian is not needed for this question and say so.
Decisive Checks
The first check is methodological. On large-angle wall cells, the potential profile across the wall from the DeepH-plus-Poisson scheme must match full OpenMX SCF to within a meV-scale tolerance. If it does not, the whole pipeline is untrustworthy and we fall back to direct linear-scaling DFT on smaller devices.
The second check is the physical headline: the zero-bias gap as a function of twist mismatch, for θ₁ near 1.08° and mismatches up to about 0.2°. We compare four curves: bare continuum, locally-uniform DFT, full calculation without screening, and full calculation with screening. The prediction to test is that the full screened gap falls well below the linear scaling of the step-potential picture.
The third check tracks how the Fano-factor peak energies and widths change with wall width and with the choice of functional. If the peaks stay pinned to each domain's singularities under hybrid exchange and screening, they remain a usable experimental probe of local twist. If they wash out, the proposal to diagnose twist disorder from noise measurements needs revising.
References
- Transport across twist angle domains in moiré graphene
- An $O(N)$ $Ab~initio$ Calculation Scheme for Large-Scale Moiré Structures
- A deep equivariant neural network approach for efficient hybrid density functional calculations
- DeepH-2: Enhancing deep-learning electronic structure via an equivariant local-coordinate transformer
- Dihedral-angle-corrected registry-dependent interlayer potential for multilayer graphene structures
- Hybrid neural network potential for multilayer graphene
- Co-Dominant Piezoelectric and Flexoelectric Effects in Twisted Double Bilayer Graphene