When a metal electrode is deposited onto an atomically thin semiconductor such as molybdenum disulfide (MoS₂), the two materials deform each other. Metal atoms push into the sheet, sulfur atoms shift or leave, and thin oxides form. Current simulations of contact resistance, the electrical penalty for getting electrons from the metal into the semiconductor, mostly ignore this. They place an ideal metal slab on an ideal sheet and compute electron transport through that frozen stack. This proposal simulates the deposition itself using a machine-learned force field. A graph neural network then predicts, atom by atom, the quantum-mechanical coupling matrix (the Hamiltonian) of the resulting disordered interface. A standard quantum transport calculation runs on that predicted Hamiltonian over contact lengths of tens of nanometers, which is far beyond what direct first-principles calculation can reach. The output is the transfer length, meaning how far under the contact electrons actually enter the semiconductor, together with whether injection happens only at the contact edge or across its whole area.
The stakes are practical. Contact resistance dominates total resistance in sub-10 nm two-dimensional transistors (The Roadmap of 2D Materials and Devices Toward Chips). Yet reported transfer lengths for MoS₂ range from about 1 nm to more than 500 nm (Electron Transport through Metal/MoS2 Interfaces: Edge- or Area-Dependent Process?), and without the transfer length nobody can say how short a contact can be made. The borrowed idea comes from drug design, where "induced fit" names the mutual reshaping of a protein and a ligand as they bind. Drug-design models that freeze the protein fail on real unbound structures. Our bet is that frozen-interface contact models fail for the same reason, and that the failure accounts for much of the transfer-length spread.
Core Claim
The injection regime and transfer length of a metal/2D contact are set by how the interface reconstructs during deposition, not by the ideal stacking alone. They can be predicted at device scale by composing three pieces. The first is a charge-aware machine-learned interatomic potential that simulates deposition. The second is an equivariant graph network that predicts local Hamiltonian blocks in a localized orbital basis (maximally localized Wannier functions). The third is a recursive Green's-function transport solve. Here "equivariant" means the network's outputs rotate consistently when the atoms are rotated.
What would be new is the link from dynamic atomic reconstruction to an electrical observable. The closest existing work, Atomistic Insights into Cu/amorphous-TaxN Interfacial Adhesion via Machine Learning Interatomic Potentials, simulates deposition at a metal interface with an equivariant potential and reaches 4.5 meV/atom energy error and 0.3 eV/Å force error. It predicts only adhesion. Its central finding, that strong intermixing is beneficial "atomic Velcro", is exactly the chemistry that is harmful for electrical contacts. We keep its simulation method and replace its target.
The transfer length has the classical form
$$L_T = \sqrt{\rho_c / R_{sh}},$$
where $\rho_c$ is the specific contact resistivity of the metal–sheet interface (resistance times area) and $R_{sh}$ is the sheet resistance of the semiconductor under the contact. Szabó et al. showed that for clean Ti/MoS₂ this picture breaks down. A contact-induced potential barrier blocks vertical injection, so current enters only at the edge. An interfacial TiO₂ layer removes that barrier and restores area injection with $L_T \approx 150$ nm (Electron Transport through Metal/MoS2 Interfaces: Edge- or Area-Dependent Process?). Both $\rho_c$ and the barrier are local, disorder-sensitive properties of the relaxed interface. That locality is what a graph model can learn from small first-principles patches and then apply to large ones.
Research Program
Stage 1: deposition ensembles. We fine-tune a universal charge-informed potential, CHGNet, on first-principles data for Ti, Ni, Au, Bi and partially oxidized Ti contacting mono- and bilayer MoS₂. CHGNet was chosen because its magnetic-moment supervision tracks oxidation state across 89 elements (CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling). Training data are grown by active learning, following the Cu/TaₓN protocol. We then run simulated vapor deposition at several incident energies and oxygen partial pressures. Each condition yields an ensemble of relaxed interfaces 30–150 nm long, the range probed experimentally. As a control, the same compositions are also built as frozen ideal stacks.
Stage 2: Hamiltonian prediction. From first-principles calculations of small interface patches, both relaxed snapshots and ideal stacks, we extract Wannier Hamiltonians. We then train an equivariant message-passing network to predict on-site and hopping blocks from local atomic environments. The design follows the finding in ∇²DFT that direct Hamiltonian prediction with PhiSNet-style equivariant networks is the strongest approach for organic molecules. Our interfaces differ from that setting in being metallic, periodic in-plane and disordered. We add CHGNet-derived site charges as node features, because charge transfer at the interface is exactly what sets pinning by metal-induced gap states (electronic states from the metal that leak into the semiconductor's band gap).
Stage 3: transport. Recursive Green's-function transport runs on the predicted Hamiltonian of the full-length contact. We compute contact resistance as a function of contact length, the position-resolved injected current, and $L_T$. Low-bias results come from linear response. For the high-bias checks, a Poisson update of the on-site energies is iterated to self-consistency.
The induced-fit lens also yields a mechanistic analysis. FlexSBDD found that local sidechain flexibility matters far more than global backbone motion for binding quality (FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling). We test the analogous split for contacts. Local reconstruction covers sulfur vacancies, metal substitution and interfacial oxygen. Global deformation covers sheet corrugation and strain. We apply each separately to the ideal stack and ask which one moves $L_T$.
Baselines
The naive reference point is a Schottky–Mott estimate: the barrier height is taken from the metal work function minus the semiconductor's electron affinity, and a fixed textbook $\rho_c$ is plugged into the transmission-line formula. The work-function inputs are the kind that monolayer databases supply and that Benchmarking graph neural networks for materials chemistry predicts on C2DB. That paper gives no number for work-function accuracy that we can quote. Because Fermi-level pinning decouples the barrier from the work function (The Roadmap of 2D Materials and Devices Toward Chips), we expect this baseline to rank metals incorrectly. Beating it means reproducing the correct ranking.
The strongest existing method is direct first-principles Wannier-basis transport on ideal commensurate stacks, as in Szabó et al. It reports edge injection for clean Ti/MoS₂ and $L_T \approx 150$ nm for Ti/TiO₂/MoS₂. Related atomistic transport work summarized in the asymmetrical-contact papers reports $L_T \approx 5$ nm for ideal Ni/MoS₂ (Contact Scaling for 2D FETs Using Asymmetrical Contact Measurements). We rerun this method ourselves on the same stacks and the same Wannier setup, so the comparison uses identical physics. Our pipeline must first reproduce these numbers on ideal stacks, which serves as a correctness check. It then has to show that it adds information on deposited interfaces, where the direct method cannot reach the needed size.
The experimental reference is the asymmetrical contact measurement: one long and one short contact on the same bilayer MoS₂ flake, with source and drain swapped between measurements. It gives $L_T < 30$ nm, where the conventional transmission-line method on the same films gives about 190 nm (Distinct Contact Scaling Effects in MoS2 Transistors Revealed with Asymmetrical Contact Measurements). Beating the frozen baseline here means that our predicted $L_T$ distribution for deposited Ni or Ti falls inside the experimental window while the frozen stack's does not.
Two ablations isolate the key ingredients. The first is our full pipeline with the Hamiltonian network applied to frozen stacks, which isolates the effect of reconstruction. The second removes the charge features from the Hamiltonian network, which isolates charge awareness. Both ablations get the same training data and the same hyperparameter budget. That budget uses multi-fidelity early stopping, following Hyperparameter Optimization of Graph Neural Networks for the OpenCatalyst Dataset.
Why not just…?
Why not just run first-principles transport on the deposited structures directly? For small patches we do exactly this, and those results are the training and test labels. At 30–150 nm contact lengths, with dozens of disorder realizations per condition, full first-principles transport is prohibitive, and the research file flags this as the field's computational limit. This alternative remains a real threat if locality fails, that is, if long-range charge rearrangement makes patch-trained Hamiltonians inaccurate at large scale. We test this directly by comparing predicted and computed Hamiltonians and transmissions on the largest patches first-principles calculation can handle, held out by size.
Why not just predict contact resistance from descriptors such as work function, or from local-environment fingerprints with kernel regression? SOAP (smooth overlap of atomic positions) is a hand-built descriptor of each atom's neighborhood, and SOAP with kernels matches graph networks on compositionally uniform systems (Benchmarking graph neural networks for materials chemistry). A single-metal study might therefore not need a graph network. We include SOAP-kernel Hamiltonian regression at matched data sizes. Our expectation is that it holds up for one metal but degrades on the held-out metals Ni and Bi, where the same study found graph networks ahead on diverse compositions. If it does not degrade, we will report SOAP as sufficient. Predicting contact resistance directly from descriptors skips the edge-versus-area physics entirely, and the Szabó clean/oxide contrast exists to test exactly that physics.
Why not use a generative flow-matching model for the induced fit, as in FlowDock or FlexSBDD? These models learn to transform an unbound structure into a bound complex in about 20 steps (FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction). For proteins that is worthwhile because sampling is expensive. For contacts, molecular dynamics with a learned potential is already cheap, and it is physically tied to deposition conditions that a generator would have to learn separately. We therefore borrow induced fit as the concept and as the key ablation, not as the sampler. Chemical-interaction work such as Representational Alignment with Chemical Induced Fit for Molecular Relational Learning likewise operates on learned representations rather than on the physical relaxation we need.
Proposed Experiments
Check 1: the ideal-stack limit. The Hamiltonian network, trained only on patches, must reproduce the direct-method results on ideal stacks. For clean Ti it should give edge injection, meaning resistance independent of contact length above a few nanometers. For Ti/TiO₂ it should give area injection with $L_T$ near 150 nm. For Ni it should give $L_T$ near 5 nm. Failure here stops the project.
Check 2: the induced-fit test. For each metal we compare $L_T$ and injection regime between frozen and deposited interfaces. The hypothesis fails if deposition changes $L_T$ by less than the spread between disorder realizations of the same condition.
Check 3: resolving the transfer-length paradox. Sweeping oxygen exposure during Ti deposition should carry the system continuously from edge to area injection. That would show the literature's 1–600 nm spread to be a chemistry variable, not only a measurement artifact. The regime classification is scored against small-patch first-principles labels.
Check 4: experimental ranking. Predicted contact resistance should reproduce Bi and Sb below Ni/Au, consistent with the roadmap's semimetal claim. It should also reproduce VS₂ contacts at roughly one quarter of Ni/Au, as measured in Metallic Vanadium Disulfide Nanosheets as a Platform Material for Multifunctional Electrode Applications. And deposited Ni/Ti should give $L_T$ below 30 nm, matching the asymmetrical-contact measurements.
Check 5: device-level asymmetry (exploratory). At high bias, with self-consistent electrostatics, we ask whether shortening the source contact to 30 nm lowers saturation current more than shortening the drain. The measured effect is a 15% drop in saturation current and a 30% drop in saturation voltage for a short source. Reproducing the qualitative source–drain asymmetry would count as success.
Across all checks we report Hamiltonian-block errors, transmission errors near the Fermi level, and the $L_T$ distribution over disorder realizations. The generalization split holds out Ni and Bi entirely from training.
References
- The Roadmap of 2D Materials and Devices Toward Chips
- Electron Transport through Metal/MoS2 Interfaces: Edge- or Area-Dependent Process?
- Atomistic Insights into Cu/amorphous-Ta$_x$N Interfacial Adhesion via Machine Learning Interatomic Potentials: Effects of Stoichiometry and Interface Construction
- CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling
- $\nabla^2$DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
- Benchmarking graph neural networks for materials chemistry
- Contact Scaling for 2D FETs Using Asymmetrical Contact Measurements
- Distinct Contact Scaling Effects in MoS2 Transistors Revealed with Asymmetrical Contact Measurements
- Hyperparameter Optimization of Graph Neural Networks for the OpenCatalyst Dataset: A Case Study
- FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction
- Representational Alignment with Chemical Induced Fit for Molecular Relational Learning
- Metallic Vanadium Disulfide Nanosheets as a Platform Material for Multifunctional Electrode Applications