Global Charge Balance as the Missing Long-Range Channel in Machine-Learned Hamiltonians for 2D Materials on Substrates

Machine-learned electronic Hamiltonians now reproduce density functional theory (DFT, the standard quantum-mechanical method for computing electronic structure) on supercells with thousands of atoms, at seconds instead of hours of cost. They work by predicting the Hamiltonian block for each atom pair from that pair's local neighborhood. A 2D material sitting on a substrate breaks this locality assumption. Electrons flow across the interface until the two materials share a single chemical potential. How much charge moves, and how far the bands shift, therefore depends on a condition that holds over the whole system: total charge neutrality and a common Fermi level. It does not depend on what any one atom sees within a few ångströms. In graph-learning terms this is the "under-reaching" failure. No number of local message-passing steps within a fixed cutoff can tell an atom what the global electron count and potential step are.

We propose to make this physical constraint the long-range channel of an equivariant Hamiltonian network. The network has two parts. A local equivariant model predicts Hamiltonian blocks, as current models do. A differentiable charge-equilibration layer adds the long-range physics. It solves, once per forward pass and under an explicit neutrality constraint, for interfacial charges that interact through a substrate-screened Coulomb kernel. It then feeds the resulting electrostatic potential back into the on-site blocks. The claim is specific and falsifiable. On monolayer–substrate systems, the dominant error of local Hamiltonian networks lies in a global, nearly rigid-plus-smooth potential component. A constrained solver captures that component more accurately and more cheaply than generic long-range graph machinery such as global attention or virtual nodes. It also transfers from small training cells to large, inhomogeneous moiré interfaces.

Where the two directions meet, and what is not yet done

The substrate literature supplies the motivation. Substrates drive spontaneous charge transfer that shifts the Fermi level of graphene nanoribbons enough to require gating to recover designed spin filtering (Dual spin filter effect in a zigzag graphene nanoribbon). Substrate bulk states hybridize with the topological edge states of a Bi(111) bilayer on Sb (Engineering Electronic Structure of a Two-Dimensional Topological Insulator Bi(111) Bilayer on Sb Nanofilms by Quantum Confinement Effect). Strain and broken mirror symmetry switch ReC2H between trivial and topological phases at about 2.3% and 7.8% biaxial strain (Switchable topological phase transition and nonlinear optical properties in a ReC2H monolayer). Freestanding models of 2D ice fail for similar reasons (Electronic and optical properties of two-dimensional ice I). These effects divide into two kinds. Hybridization and strain are local and fit within a cutoff. Charge transfer and its electrostatics are global.

The closest existing work is HamGNN (Transferable equivariant graph neural networks for the Hamiltonians of molecules and solids). It is an equivariant network that fits DFT Hamiltonians with spin–orbit coupling. It reaches about 1.5 meV mean absolute error (MAE) on matrix elements and computes a 1,626-atom twisted MoS2 Hamiltonian in 21 seconds versus about 20 hours for DFT. Its long-range ability comes from equivariant pair features up to a radial cutoff, and all its large-cell tests are freestanding supercells in vacuum. Our proposal reuses its local machinery. What we add is a mechanism for information that no cutoff contains.

The graph-learning literature supplies both candidate tools and a warning. The proposed solutions are global attention, as in the GPS hybrid of local message passing and linear attention (Recipe for a General, Powerful, Scalable Graph Transformer); virtual nodes (Manolache et al.); error-driven rewiring (Error-Driven Graph Augmentation for Mesh-Based PDE Surrogates); and non-dissipative dynamics (Improving Long-Range Interactions in Graph Neural Simulators via Hamiltonian Dynamics).

Can You Hear Me Now? A Benchmark for Long-Range Graph Propagation found that DFT partial charges on molecules (ECHO-Charge) require such machinery. This motivates our work, because charges are exactly the quantity that sets the long-range potential. Where Did the Gap Go? Reassessing the Long-Range Graph Benchmark showed that tuned plain message passing matches graph transformers on many "long-range" tasks. This constrains our work: we must show that the long-range dependency is real in our data before crediting any architecture.

None of these papers studies periodic interfaces or substrate-induced charge transfer. None of the long-range graph methods produces equivariant Hamiltonians. The obvious combination, attaching GPS-style attention to HamGNN, would be new on paper. We think it is the wrong design. Attention must learn from data a constraint (neutrality plus a screened Coulomb law) that we can write down exactly. It also scales poorly on thousand-atom cells.

Core Claim

Write the predicted Hamiltonian as a local part plus an electrostatic correction on the diagonal:

$$H_{ij} = H^{\text{loc}}_{ij}(\mathcal{N}_{ij}) + \delta_{ij}\, V_i \, S_{ii},\qquad V_i = \sum_j K_\epsilon(r_i, r_j)\, q_j .$$

Here $H^{\text{loc}}_{ij}$ is the equivariant block predicted from the local neighborhood $\mathcal{N}_{ij}$. $S_{ii}$ is the on-site overlap block. $V_i$ is the electrostatic potential at atom $i$. $q_j$ are atomic charges. $K_\epsilon$ is a periodic Coulomb kernel screened by a substrate dielectric parameter $\epsilon$, using an image-charge form for a half-space dielectric.

The charges come from a differentiable equilibration step. The network predicts local electronegativities $\chi_i$ and hardnesses $\eta_i$. The step then minimizes $\sum_i (\chi_i q_i + \tfrac12 \eta_i q_i^2) + \tfrac12 \sum_{ij} q_i K_\epsilon(r_i,r_j) q_j$ subject to $\sum_i q_i = Q$, where $Q$ is the total charge. This is a single linear solve with a Lagrange multiplier. That multiplier is the global chemical potential, the quantity a local model cannot represent. We solve it with Ewald summation or an iterative method at roughly $O(N\log N)$ cost.

The claim has three parts. First, a local model's error on substrate systems is concentrated in a component that varies smoothly across the cell and is set by global conditions. Second, the equilibration layer removes most of that error. Third, it does so better than attention, virtual nodes, or a larger cutoff at matched compute. The gain should be largest on large, laterally inhomogeneous interfaces such as moiré patterns and charge puddles, where a single rigid shift is insufficient.

Research Program

The first step is building data, because the file contains no substrate-interface benchmark. We will compute DFT Hamiltonians in a localized-orbital basis, as HamGNN uses, for five families:

  1. graphene and zigzag nanoribbons on metal slabs;
  2. the Bi(111) bilayer on Sb films of varying thickness;
  3. monolayer NbSe2 on SiO2-like insulators and on graphene;
  4. ReC2H on substrates that impose strain across the 0–9% range spanning both topological transitions;
  5. 2D ice on a polar surface.

Training cells stay below about 60 atoms, matching HamGNN's protocol. Test sets contain larger moiré or mismatched supercells of 1,000–3,000 atoms where DFT is still affordable, plus thickness and lateral-size extrapolations.

The second step is a controlled diagnostic of range, built before any architectural comparison. We construct paired structures that are identical within the model cutoff around every monolayer atom but differ globally. The pairs vary in substrate slab thickness, include a distant dopant layer, or use a different substrate termination several nanometres away. We then measure how much the DFT Hamiltonian differs within each pair. Any cutoff model must predict identical outputs for both members of a pair, so the DFT difference is a hard lower bound on its error. This gives Gap 3 of the file a physical answer rather than a benchmark-artifact answer. If the differences come out below roughly the 1.5 meV accuracy floor, local models suffice and the proposal's premise is wrong.

The third step builds the model. We start from a HamGNN-style equivariant backbone and add the equilibration layer. For the substrate dielectric, we compare treating $\epsilon$ as an input against treating it as a learned per-substrate parameter.

The fourth step examines the physics downstream of the Hamiltonian. This covers the Fermi-level shift, charge transfer, gap and band-inversion location, the Z2 invariant along the ReC2H strain path, and whether the Bi bilayer's edge states survive on Sb.

Many-body screening beyond DFT, such as the GW effects that move the NbSe2 Van Hove singularity from −123 to −77 meV (Quasiparticle energy bands and Fermi surfaces of monolayer NbSe2), is out of scope as a training target, since our labels are DFT. We note it as the natural next target, once the electrostatic channel exists and can carry a dielectric parameter.

Baselines

The naive reference point is a HamGNN-style model trained on freestanding monolayers and applied to the supported system with no substrate atoms. This measures how wrong the vacuum abstraction is.

The strongest existing method is HamGNN itself, trained on the substrate cells with substrate atoms included and its cutoff tuned. Its reported figure is about 1.5 meV MAE on Hamiltonian elements for its own freestanding datasets. We found no reported result on Lacuna for HamGNN or any machine-learned Hamiltonian on monolayer–substrate interfaces, so we will rerun it ourselves.

Fairness rules apply to all trained models. They share the same training cells and the same equivariant backbone width. Each receives an equal Bayesian-optimization budget, as ECHO uses, and the normalization and output-head practices that Where Did the Gap Go? showed matter. Comparisons are reported at matched GPU-hours.

Three generic long-range competitors are attached to the same backbone. One is GPS-style linear attention (Recipe for a General, Powerful, Scalable Graph Transformer). One is a virtual node (Manolache et al.). One is a non-dissipative deep propagation stack in the spirit of the models that led ECHO-Charge, where SWAN reports MAE 6.109×10⁻³ (Can You Hear Me Now?).

A physics baseline applies a post-hoc rigid Fermi alignment plus a fixed analytic dipole correction to the local model.

The key ablation removes the neutrality constraint, leaving the screened Coulomb layer with unconstrained charges. This isolates whether the global multiplier, rather than merely having a Coulomb kernel, carries the gain.

Beating the baselines means two things. The model must cut large-cell MAE and Fermi-shift error relative to tuned HamGNN by more than seed variance. It must also match or beat attention at lower cost, on held-out larger cells and in the thickness extrapolation.

Why not just…?

Why not just tune a local equivariant model with a larger cutoff and better optimization? This is the lesson of Where Did the Gap Go?, and we take it seriously. The paired-structure diagnostic answers it directly. If DFT Hamiltonians differ within pairs that are identical inside the cutoff, no tuning can close that gap. The size of the difference bounds how much a larger cutoff can buy. If the difference is negligible, we will report that and drop the long-range claim.

Why not just add global attention or a virtual node? Both can in principle learn a global chemical potential. A virtual node is a particularly natural place to store it, and this makes the virtual node a real competitor. Its weakness is that it must learn neutrality and the screened Coulomb law from small training cells. Our expectation is that it then extrapolates poorly to larger, laterally inhomogeneous cells. The cost gap also matters: in the ECHO benchmark GPS trained at about 788 s/epoch on ECHO-Charge versus about 12.8 s/epoch for A-DGN. The large-cell transfer test will settle this question. If a virtual node matches us there, the honest conclusion is that a single global scalar suffices. The constrained solver would then be justified only for inhomogeneous moiré cases.

Why not just apply a rigid Fermi-level shift and an analytic dipole correction after a local prediction? For laterally uniform interfaces this may work nearly as well as our model, so it is the most serious threat. Our model differs in two ways. It produces spatially varying potentials, which matter for moiré charge modulation and for lattice-mismatched cells. And the charges are predicted self-consistently from local chemistry rather than fixed in advance. We will report results separately for uniform and inhomogeneous test cells. The claim survives only if the gap to this baseline grows with lateral inhomogeneity.

Proposed Experiments

The headline experiment trains all models on small supported cells and tests on 1,000–3,000-atom mismatched and moiré interfaces. It reports four quantities: Hamiltonian MAE stratified by distance from the interface, the error in the Fermi-level shift relative to the freestanding layer, the error in interfacial charge transfer, and band-structure fidelity near the Fermi level.

A thickness-extrapolation experiment trains on Sb films of a few layers and tests on thicker films. It asks whether the Bi bilayer's edge-state hybridization and gap are reproduced.

A phase-diagram experiment predicts the Z2 invariant and band-inversion location along the ReC2H strain path on a substrate. Success means recovering transitions near the reported 2.3% and 7.8% values, or their substrate-shifted counterparts in our DFT, with the correct inversion momenta.

Two sanity checks test the long-range layer outside the substrate setting. The first places the equilibration layer, without the Hamiltonian head, on ECHO-Charge, to check that it is competitive with SWAN's reported 6.109×10⁻³ MAE on a known long-range charge task. The second reports wall-clock cost against GPS at matched accuracy.

The decisive check is the paired-structure diagnostic combined with the rigid-shift baseline on inhomogeneous cells. Together they determine whether substrate electronic structure carries a genuinely global dependency, and whether a constrained physical solver is the right way to transmit it.