Deciding When Domain-Wall Details Matter for Moiré Transport

Twisted graphene samples are never uniform. The angle between the layers drifts across the flake, splitting it into patches separated by thin boundaries called domain walls. A useful simplification treats each wall as a "step potential": a sudden jump in energy whose height grows with the angle mismatch on either side. Electrons crossing a chain of such walls then behave like waves passing a stack of steps. These same walls are also electromechanically active, meaning they deform when a voltage or a probe force is applied. They have measured piezoelectric and flexoelectric responses, so their shape, and hence their potential profile, depends on bias and local strain.

This raises a practical question that current tools cannot answer exhaustively. Suppose we refine the simple step model into a smooth, strained wall whose shape responds to voltage. Does the refinement change predicted transport for some arrangement of walls, or for none? Today the answer comes from simulating a handful of configurations. We propose to answer it exactly. A chain of walls and domains is a product of transfer matrices, which makes it a weighted automaton: a finite-state machine that assigns a number to every input sequence. Here the input sequence lists the walls and domains, and the number is the transmission. Two wall models can then be compared for every possible disordered sequence at once using the polynomial-identity-testing equivalence checks developed for probabilistic programs. When the models differ, the checker returns the shortest wall sequence that tells them apart, which is a concrete device design for an experiment.

Core Claim

The claim is that the "open component in an arbitrary context" view from algorithmic game semantics transfers to coherent moiré transport. In that view, two components are interchangeable exactly when no surrounding context can tell them apart. Here the component is one domain-wall model. The context is everything else in the sample: the other walls, the domains between them, and the leads. The context is chosen by an adversary, which in game-semantic terms is the Opponent, while the device's response is the Player. Contextual equivalence of two wall models then becomes language equivalence of weighted automata.

One step makes this algorithmically sharp. If the propagation phase across each domain is treated as a formal variable rather than a fixed number, then "equivalent for all domain lengths" becomes a polynomial identity. That is exactly the setting where Schwartz–Zippel random evaluation is sound. What is new is not faster simulation. It is a decision procedure, plus a minimal witness, for the question of which electromechanical details of a wall can ever become visible in conductance.

Grounding in prior work

The physical starting point is Transport across twist angle domains in moiré graphene. That work maps twist-angle walls onto step potentials of height proportional to $|\theta_1-\theta_2|$ in a two-band model. It computes transmission with transfer matrices and obtains a zero-bias transport gap and sub-Poissonian Fano-factor peaks near van Hove energies. It supplies our transfer matrices and our validation targets. It also motivates the work, because it explicitly idealizes walls as sharp, flat, and semi-infinite. Whether that idealization is safe is precisely the question we make decidable.

The electromechanical refinements come from three sources. The first is Layer-Dependent Electromechanical Response in Twisted Graphene Moiré Superlattices. It reports purely flexoelectric walls in marginally twisted bilayer graphene (0.05 pm/V) and piezoelectric response in twisted monolayer-bilayer graphene (0.35 pm/V), with more than tenfold higher sensitivity to tip force. The second is Co-Dominant Piezoelectric and Flexoelectric Effects in Twisted Double Bilayer Graphene. It reports a two-step "piezo-flexoelectric" wall profile in twisted double bilayer graphene (0.15 pm/V), reproduced by continuum simulation. These papers parameterize our refined wall letters. They treat walls only as static transducers and never feed the deformation back into transport. Coupling the two is what we do.

The third source is Are better conductors more rigid?, which shows that local carrier occupancy changes stiffness, through what it calls electro-stiffness, scaling as $t^2/\sqrt{u^2+t^2}$. It motivates a bias-dependent wall whose shape depends on the charge piled up by reflection. It is a 1D chain result, so we use it only as a qualitative template for a feedback term with a free coupling constant, not as a quantitative wall model.

The algorithms come from Direction 3. Language Equivalence for Probabilistic Automata and On the Complexity of the Equivalence Problem for Probabilistic Automata reduce equivalence to linear algebra and polynomial identity testing. The first reports an $O(n^3)$ deterministic method, an $O(n\cdot|M|)$ randomized one, and counterexample extraction at under 12% overhead. Neither relies on stochasticity, so both apply to weighted automata over the complex numbers. APEX: An Analyzer for Open Probabilistic Programs and Algorithmic probabilistic game semantics - Playing games with automata supply two things: the notion of an open component checked against all contexts, and bottom-up compilation with state reduction. Their restriction to finite ground types acts as a constraint. It is why angles and wall widths are binned into a finite alphabet, while domain lengths are handled symbolically.

Research Program

The construction proceeds in four steps.

Building the alphabet. We restrict to parallel, flat walls, so the transverse momentum $k_y$ is conserved and each channel can be treated separately. Within one channel, the alphabet contains two kinds of letters. Domain letters are indexed by a binned twist angle. Wall letters are indexed by an angle-pair bin and a wall model: sharp step; smooth finite width; strained with a flexoelectric profile; strained with a two-step piezo-flexoelectric profile; or bias-dependent with feedback.

Defining the weight. Each letter $a$ has a transfer matrix $M_a(E,k_y)$, where $E$ is energy. A domain letter carries a free phase variable $z_a$ in place of $e^{ik\ell}$. The transmission amplitude follows from the entry $(M_w)_{22}$ of the word's product matrix $M_w$. Equality of $|t|^2$ is therefore equality of

$$f(w)=\langle e,\,(M_w\otimes \overline{M_w})\,e\rangle,$$

where $w$ is a wall/domain sequence, $\otimes$ is the Kronecker product, the bar denotes the complex conjugate, and $e$ selects the $(2,2)$ entry. In the conjugate copy, $\bar z_a$ is replaced by $z_a^{-1}$, so $f$ is a Laurent polynomial in the phases and remains linear in the automaton sense. We will verify that this substitution is correct.

Adding context dependence. Electrostatic coupling between neighboring walls is represented by automaton states that remember the last $k$ letters. A wall's matrix then depends on its context. This is where genuine automaton structure appears, rather than a bare matrix product.

Running the checks. Equivalence is tested with the randomized check: random rational phases and random letter weights, together with a deterministic linear-algebra cross-check. The checks run per $(E,k_y)$ on a grid, using exact rational or interval arithmetic for the polynomial evaluations. That choice addresses the floating-point drift noted in Language Equivalence for Probabilistic Automata.

Scope limits. Exact equivalence is a sharp yes/no answer. Real refinements will almost always be inequivalent, so the more useful outputs are the shortest distinguishing word and its discrepancy, and a ranking of which refinements first become visible at which sequence length. Searching for the maximum discrepancy over all words is not decidable in general. We therefore do not claim worst-case error bounds, only exact verdicts plus guided witnesses. Tilted walls that mix channels give only a sufficient condition, through matrix-block equality, and are reported separately.

Baselines

The do-nothing reference is the sharp-step model of Transport across twist angle domains in moiré graphene, taken as adequate. Every refined model is compared against it on the transport-gap and Fano-peak observables. If the checker finds the refinements equivalent to it on all channels in our energy window, the idealization is vindicated, which is itself a publishable result. No reported numerical Fano or gap value was found on Lacuna for this benchmark, so agreement is judged against our own reproduction of its qualitative trends: monotone gap scaling with angle mismatch, and peaks at the van Hove energies.

The strongest practical competitor is random-sequence Monte Carlo. It draws disordered wall/domain sequences, computes conductance and Fano factor for both models, and compares the distributions. This is how disorder is characterized in Percolation and Transport Properties in The Mechanically Deformed Composites Filled with Carbon Nanotubes. It is cheap and directly physical. It gets the same alphabet, the same matrices, the same energy grid, and the same wall-clock budget as the checker. The proposal beats it if one of two things holds. Either the checker finds distinguishing sequences that Monte Carlo misses within its budget, or the checker certifies equivalences that Monte Carlo can only fail to refute.

Bounded brute-force enumeration of every word up to length $n_1+n_2$ is a second exact competitor. Any two inequivalent automata with $n_1$ and $n_2$ states differ on some word of that length, so this enumeration is complete in principle, but it costs $|\Sigma|^{n_1+n_2}$, where $|\Sigma|$ is the alphabet size. For the solver itself, deterministic QR-based equivalence is compared against randomized testing. The only reported figures are from software benchmarks: on Herman's protocol with $N=15$, the randomized method took about 4 seconds against 93 seconds for the optimized deterministic one (Language Equivalence for Probabilistic Automata). We treat these as a sanity reference, not a target, because our automata are much smaller.

Two ablations isolate the key ingredients. The first fixes domain lengths numerically instead of keeping symbolic phases. It tests whether "equivalence for all lengths" changes verdicts relative to sampled lengths. The second removes the context-memory states. It tests whether neighbor-dependent feedback is ever what makes two models distinguishable.

Why not just…?

Why not just compare single-wall transmission of the two models? In the incoherent regime this is the right answer, and we concede it. Resistances of walls then add classically, and single-wall transmission determines everything. In the coherent regime it fails in both directions. Two walls with identical $|t|^2$ but different reflection phases produce different interference across a stack. Walls whose matrices differ by a transformation invisible at the leads remain indistinguishable even though their matrices differ. Our planted-pair experiment, described below, makes both failure modes explicit. If phase-averaging over domain lengths erased every difference beyond the single-wall level, the symbolic-phase ablation would show it, and we would report that the automaton adds nothing.

Why not just run Monte Carlo over disorder? This is the real threat. If the differences between refinements are generic, random sequences will expose them almost immediately, and the checker's main remaining value is certified equivalence plus a minimal witness. We measure directly how many random samples are needed to reach the discrepancy that the checker's shortest witness achieves. We will report honestly if that number is small.

Why not use a closed probabilistic model checker like PRISM? The quantity of interest is a complex amplitude, not a probability. More fundamentally, we need equivalence over all contexts, not a property of one closed system. APEX: An Analyzer for Open Probabilistic Programs reports that closed checkers fail beyond double-digit participants on open protocols, while compositional checking reaches 800. Our automata are small, though, so the argument here rests on the open-context semantics rather than on scale.

Proposed Experiments

Reproduction. The first check reproduces the qualitative zero-bias gap and Fano-peak behavior of Transport across twist angle domains in moiré graphene from the automaton representation.

Planted pairs. The second check uses planted pairs: gauge-transformed walls that should be equivalent, and phase-shifted walls with equal $|t|^2$ that should be inequivalent. The checker must classify every pair correctly with exact arithmetic, and its witness lengths must respect the $n_1+n_2$ bound.

Main study. The main study builds an equivalence and witness map across the five wall models, over energies spanning both van Hove windows. Electromechanical profiles are scaled to the reported 0.05, 0.15 and 0.35 pm/V responses, and strain is computed for tip forces in the range used by DART LPFM. Results are reported per stacking type. The decisive outcome is a ranking that says, for example, whether the two-step piezo-flexoelectric wall of twisted double bilayer graphene becomes visible in conductance at a shorter wall sequence than the purely flexoelectric bilayer wall. That ranking is a prediction checkable in a gated device whose domain sequence is mapped by piezoresponse imaging.

Head-to-head comparisons. We report witness discovery rate against Monte Carlo and brute force at matched budgets, and both ablations. The project is judged a failure if every refinement is detectable by random sampling within a few dozen draws and no non-trivial equivalences are found. In that case the honest conclusion is that simulation suffices.