Theory predicts that a stripe-shaped electric potential, repeating every few nanometres across graphene, can bend the material's electronic band structure so that electrons keep moving at full speed along the stripes but almost stop moving across them. This effect, called anisotropic velocity renormalization, would let a sheet of graphene steer electron beams with no walls or waveguides. Simulations show beams staying collimated to 0.3° over 0.1 mm (Electron Beam Supercollimation in Graphene Superlattices). The difficulty is building the potential. The room-temperature regime needs roughly a 10 nm period and an amplitude near 0.5 eV (New Generation of Massless Dirac Fermions in Graphene under External Periodic Potentials). Lithographic gates cannot draw features that fine. Epitaxial overlayers do produce anisotropy, but only weakly, and in ways density functional theory fails to predict (In-plane anisotropy of graphene by strong interlayer interactions with van der Waals epitaxially grown MoO3).
We propose a different source for the potential: the alternating "up" and "down" polarization stripes that form on their own in thin ferroelectric oxide films grown by pulsed laser deposition (PLD). PLD fires laser pulses at a solid target and deposits the vaporized material one atomic layer at a time. The stripe period in such films is set by film thickness and strain, and it naturally falls in the nanometre range. We pair this physical route with a staged design tool. It is modelled on multi-level intermediate representations (MLIR), the compiler practice of lowering a program through several explicit levels of description rather than jumping straight to machine code. Here the tool turns a requested electron-velocity ratio into a growth recipe. The middle level, the potential profile, is something we can measure, so every prediction the tool makes can be checked partway through.
Core Claim
The claim has two parts, and both can be tested.
The physical part: graphene transferred onto a PLD-grown ferroelectric film with stripe domains can reach a larger in-plane conductance anisotropy than the 1.43 reported for graphene on MoO3. It should also show the band-structure fingerprint of new Dirac points that only a periodic potential produces.
The methodological part: splitting the design problem into explicit levels finds recipes that hit a target anisotropy in fewer growth runs than black-box optimization over raw recipe parameters. Those levels are, from top to bottom, the target velocity tensor, the continuum potential, the film structure, and the PLD recipe. The advantage comes from two sources. Each level can be checked against an affordable measurement. And the potential level records disorder statistics explicitly rather than assuming an ideal waveform.
The second part is a real hypothesis and could be false. The experiments are designed so that a negative result is informative.
Why This Is Not Already Done
The closest existing work falls into three groups.
Graphene on an epitaxial overlayer. The MoO3 study grows an oxide on top of graphene by van der Waals epitaxy and measures the anisotropy that results. It does not design the period or amplitude. Its own DFT model underestimates the effect, which is evidence that a calibrated, measurement-anchored model is needed rather than a purely first-principles one.
PLD-grown iridate superlattices. Reconciling Monolayer and Bilayer J_eff=1/2 Square Lattices in Hybrid Oxide Superlattice and Magnetism in iridate heterostructures leveraged by structural distortions show that PLD layer sequences control octahedral tilts (about 8°) and magnetic anisotropy. Their recipes come from empirical calibration. They address spin anisotropy, not carrier velocity.
Multi-level compilers for hardware. ScaleHLS, COMET, and CIM-MLC show that keeping several explicit levels lets a search engine apply optimizations a flat representation hides. All of them target digital or crossbar hardware. In all of them, the level-to-level lowering is exact by construction.
What is new here has two pieces. First, the physical route: ferroelectric stripe domains used as the periodic potential for velocity renormalization. Second, a lowering chain in which each step is an uncertain, calibrated physical model with its own measurable checkpoint. That second feature does not exist in any of the compiler work above.
Research Program
The tool has four levels.
Target level. This states the desired ratio of perpendicular to parallel velocity, $v_y/v_x$, and the energy window that must contain only the renormalized states.
Potential level. This describes the stripe potential seen by graphene: period, amplitude, harmonic content (sinusoidal versus square-wave), and the statistics of domain-wall roughness and period scatter. The step from the target level to this level is essentially analytic. Park et al.'s continuum analysis ties collimation to a specific combination of amplitude and period. Landau Levels and Quantum Hall Effect in Graphene Superlattices gives a disorder tolerance of about 10%, which becomes a check the tool enforces before passing a design down to the next level.
Film-structure level. This describes ferroelectric layer thickness, superlattice repeat (for example $n$ ferroelectric to $m$ paraelectric unit cells), and substrate-imposed strain. The step from the potential level to this level uses domain electrostatics. The main relation is Kittel scaling:
$$w = \sqrt{\kappa\, d},$$
Here $w$ is the stripe width, $d$ is the film thickness, and $\kappa$ is a material constant fitted from our own films. The square-root form comes from balancing the energy cost of domain walls against the electrostatic energy of the depolarizing field. The amplitude is set by polarization after screening by graphene's own carriers. This screening is the main physical risk: it may cap the amplitude below 0.5 eV. The tool must flag such a target as unreachable rather than silently degrade it.
Recipe level. This holds laser pulse counts per target, fluence, substrate temperature, and oxygen pressure. Thickness is calibrated layer by layer during growth.
We will implement the levels as MLIR-style typed dialects. The reason is concrete: each level gets a verifier, so a step that produces an infeasible design fails loudly. The scientific claim, however, concerns the level structure, not the software, and the ablations below test that structure directly.
Each level is paired with a measurement, and each measurement becomes training data for one lowering model. Film-structure predictions are checked by X-ray diffraction and piezoresponse force microscopy of the bare film. Potential-level predictions are checked by Kelvin-probe or scanning-tunnelling maps of the transferred graphene. Target-level predictions are checked by angle-resolved four-probe transport and magnetotransport. Search proceeds level by level, in the spirit of ScaleHLS's neighbour traversal: candidates are proposed at the potential level and lowered downward. Expensive growth runs are spent only on candidates whose upper-level predictions are confident. Each measured film updates only the model whose checkpoint it informs.
Baselines
Naive reference. The do-nothing point is graphene transferred onto a PLD-grown, non-polar SrTiO3 film of matched thickness and roughness, processed identically. It should show near-unity conductance anisotropy. Without a potential, transport is isotropic, the case Park et al. associate with 55° beam spreading. Every anisotropy we claim is measured relative to this control on the same chip-fabrication flow.
Strongest existing physical result. The strongest comparable measured result is graphene on van der Waals-epitaxial MoO3, which reports an in-plane conductance ratio of 1.43 at $V_G = 0$ V with mobility of 8155 cm² V⁻¹ s⁻¹ (In-plane anisotropy of graphene…MoO3). It is the right competitor because it is the only measured superlattice-induced anisotropy in graphene in the file. We will reproduce it on our own transfer and measurement line, so that differences in contacts and geometry do not favour either side. Beating it means a ratio above 1.43, measured on the same device geometry, at mobility within a factor of two of 8155 cm² V⁻¹ s⁻¹. That mobility condition rules out a "win" that comes merely from damaging one direction.
Strongest search baseline. This is flat optimization over raw recipe parameters with the measured anisotropy as the objective. We will run both Bayesian optimization and a genetic algorithm of the type COMPASS uses for partitioning (COMPASS). No reported result found on Lacuna for automated recipe search in PLD, so there is no published number to beat. The comparison is internal. Every method gets the same number of growth runs, the same candidate recipe ranges, the same measurements, and the same hyperparameter-tuning budget in simulation. Beating it means reaching a target anisotropy within a stated tolerance in reliably fewer growths across repeated runs.
Ablation of the key ingredient. The two ablations isolate what the proposal adds. The first keeps the full chain but uses an ideal, disorder-free potential level. The second deletes the potential level and learns a direct film-structure-to-transport model.
Why not just…?
Why not just run Bayesian optimization directly on the PLD parameters? This is the most serious threat. With a recipe space of five or six continuous variables, flat optimization may need only a few dozen growths. Compiler results such as ScaleHLS's 768.1× kernel speedups come from exact lowering and cannot be assumed to transfer to noisy physics. We settle the question in two ways. The first is a simulated testbed in which the true recipe-to-transport map contains hidden nonidealities: interdiffusion, wall roughness, and screening. The second is a matched-budget real campaign. If flat optimization matches the staged search in growth count, we will report that. The remaining value of the levels would then be diagnosis: knowing which physical step failed, which flat optimization cannot provide.
Why not just use lithographic gate arrays? Gates give clean, tunable potentials at periods of tens of nanometres. Park et al.'s estimates, however, put the room-temperature window (about 82 meV above the 25 meV thermal scale) at around 10 nm periods, where gate fringe fields smear the potential. We include a gate-array device at the smallest pitch we can fabricate, giving a direct, same-graphene comparison of achievable anisotropy versus period.
Why not just use an intrinsically anisotropic crystal? Predicted mobility ratios above 600:1 in MnNX-type monolayers (Wang et al.) far exceed anything we expect, so on raw ratio this alternative may win. What it cannot do is offer a tunable period or amplitude, or produce the new-Dirac-point fingerprint. Our claim is controllability and design accuracy, not record anisotropy. We will state this restriction plainly.
Proposed Experiments
Stage one is the simulated testbed. A hidden ground-truth chain combines continuum Dirac transport with randomized domain disorder. All search methods are compared at fixed growth budgets, and we report growths-to-target and final error.
Stage two grows ferroelectric/paraelectric oxide superlattices by PLD at three or more thicknesses. This fits $\kappa$ in the Kittel relation and checks the film-structure level against piezoresponse force microscopy.
Stage three transfers graphene onto these films and measures three things: angle-resolved conductance anisotropy against the naive control and the MoO3 reproduction; potential maps against the potential-level predictions; and magnetotransport.
The decisive checks are threefold. First, the measured anisotropy must exceed 1.43 at preserved mobility. Second, the predicted-versus-measured anisotropy error on held-out recipes must be smaller than the run-to-run spread of the naive control. Third, Landau-level degeneracy steps of $4(2N+1)e^2/h$ must appear (Landau Levels and Quantum Hall Effect in Graphene Superlattices). This last signature identifies new Dirac points from a periodic potential and cannot be produced by uniaxial strain or doping alone. If anisotropy rises but the fingerprint never appears, we conclude that the effect is strain-driven, as in the MoO3 case, rather than true velocity renormalization.
References
- Electron Beam Supercollimation in Graphene Superlattices
- New Generation of Massless Dirac Fermions in Graphene under External Periodic Potentials
- In-plane anisotropy of graphene by strong interlayer interactions with van der Waals epitaxially grown MoO 3
- Reconciling Monolayer and Bilayer <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>J</mml:mi></mml:mrow><mml:mrow><mml:mi>eff</mml:mi></mml:mrow></mml:msub><mml:mo>=</mml:mo><mml:mn>1</mml:mn><mml:mo>/</mml:mo><mml:mn>2</mml:mn></mml:mrow></mml:math> Square Lattices in Hybrid Oxide Superlattice
- Magnetism in iridate heterostructures leveraged by structural distortions
- ScaleHLS: A New Scalable High-Level Synthesis Framework on Multi-Level Intermediate Representation
- COMET: A Domain-Specific Compilation of High-Performance Computational Chemistry
- CIM-MLC: A Multi-level Compilation Stack for Computing-In-Memory Accelerators
- Landau Levels and Quantum Hall Effect in Graphene Superlattices
- COMPASS: A Compiler Framework for Resource-Constrained Crossbar-Array Based In-Memory Deep Learning Accelerators
- A family of high-temperature ferromagnetic monolayers with locked spin-dichroism-mobility anisotropy: MnNX and CrCX (X = Cl, Br, I; C = S, Se, Te)