Some engineered quantum devices carry current along one-dimensional "valley channels". A valley is one of two equivalent momentum pockets in graphene's electronic structure, and in these channels electrons moving forward and electrons moving backward sit in opposite valleys. An electron can only bounce backward if something sharp, such as an atomic vacancy, kicks it from one valley to the other. That process is called intervalley scattering. In principle, then, any resistance above the ideal quantized value is a direct readout of intervalley scattering. In practice, the one recent device that realized such channels in bilayer graphene measured a resistance slightly above the quantum limit and attributed the excess to intervalley scattering by atomic defects. It never showed how much of the excess comes from contacts, the bent gate geometry, strain, or leakage through the bulk (Topological Valley Transport in Bilayer Graphene Induced by Interlayer Sliding). This is a residual-systematics problem of exactly the kind precision neutrino physics has learned to solve, and the tools that solved it there are largely design tools rather than machine-learning tools.
We propose to turn that qualitative attribution into a measurement with a stated systematic error budget. Three ideas carry over from reactor-neutrino practice, and they fit together. The first is a within-device length ratio: voltage probes at several spacings along the same channel, analogous to comparing detectors at different baselines. The second is a set of orthogonal cross-checks in which the same scattering length is extracted from knobs (displacement field, temperature, deliberate defect dose) that each expose a different set of confounders. The third is a frozen, blinded analysis validated for frequentist coverage on a synthetic benchmark before any real data are unblinded. The deliverable is an intervalley mean free path with a calibrated 68% interval, plus a decomposition of the excess resistance into its sources.
Core Claim
The claim rests on the symmetry of the channel. Backscattering in a valley channel requires valley transfer, so the part of the excess resistance that grows with channel length measures intervalley scattering alone. Contacts, probe interfaces and quantum reflection at the channel ends do not grow with length. In the simplest Landauer picture of $N=8$ channels with transmission $\lambda_{iv}/(\lambda_{iv}+L)$, the resistance between probes separated by length $L$ is
$$ R(L) = \frac{h}{8e^2}\left(1 + \frac{L}{\lambda_{iv}}\right) + R_s(D, T), $$
where $\lambda_{iv}$ is the intervalley mean free path and $R_s$ collects every length-independent series contribution. $R_s$ is allowed to depend on displacement field $D$ and temperature $T$. A single-length measurement cannot separate $\lambda_{iv}$ from $R_s$, and that degeneracy is why the attribution in the sliding-channel paper stays qualitative.
The length lever by itself is a standard technique, the transfer-length method. What is new is the full systematics treatment around it. Other mechanisms also scale with length: segments where the domain-wall gap closes, thermally activated leakage into the bulk, and probe invasiveness. Each of these responds differently to $D$, to $T$, and to deliberately added short-range defects. Requiring that the same $\lambda_{iv}$ come out of every knob is our analog of Daya Bay's independent measurement through the hydrogen-capture channel. That analysis validated $\theta_{13}$ through a channel with an entirely different background and calibration budget, and found inter-detector event ratios agreeing within 0.6% (Independent measurement of the neutrino mixing angle θ13 via neutron capture on hydrogen at Daya Bay).
Research Program
The program starts in simulation, because the methodology has to be shown to give correct coverage before it can be trusted on scarce devices. We will build a tight-binding model of bilayer graphene with an AB-to-BA stacking domain wall, the structure created by draping the bilayer over a nano-ridge. Into it we will place atomic vacancies at controlled density, which sets the true $\lambda_{iv}$. We will then add nuisance effects drawn from known priors: contact and probe interface resistance, a smooth electrostatic inhomogeneity across the ridge, strain-modified hopping near the ridge edges, long-range charge puddles, and a thermally activated bulk leakage path.
Smooth potentials carry too little momentum to cross between valleys. They should therefore change $R_s$ and the leakage path but not the slope in $L$. The simulation will test whether that holds in a finite device, and where it breaks down, for example when puddles close the gap locally. This plays the same role as the "known unknowns" benchmarks of FAIR Universe – the challenge of handling uncertainties in fundamental science and FAIR Universe HiggsML Uncertainty Dataset and Competition. Those benchmarks inject parameterized distortions into collider simulations and score interval estimates with a penalty that punishes under-coverage. We borrow their evaluation logic but not their collider data or their neural estimators, since our nuisance space is roughly six-dimensional and an explicit profile likelihood remains tractable. We will also include one non-parametric distortion that the inference model never sees, a random local gap-closing segment, to probe the "unknown unknowns" failure mode those benchmarks leave open.
The inference model is a joint likelihood over all probe spacings, displacement fields and temperatures for one device. It has a single shared $\lambda_{iv}$ per defect state, and nuisance terms whose functional forms are motivated physically. Bulk leakage is taken as activated in $T$ and suppressed as the gap grows with $D$. Series resistance is length-independent but may vary with $D$. Probe invasiveness is modeled as a per-probe partial-equilibration parameter. Each knob tests a different weakness. Changing $D$ separates gap-closing and leakage effects, which depend strongly on $D$, from defect-driven intervalley scattering, which should be nearly flat once the gap is open. Changing $T$ separates activated leakage from low-temperature defect scattering. The 80 K persistence reported for the ballistic channel bounds the window we use.
The third knob is a deliberate defect dose, the analog of a calibration source. Short-range scatterers are added in steps while contacts and geometry stay fixed. The predicted signature is that $1/\lambda_{iv}$ rises linearly with dose while the intercept $R_s$ does not move. Yanık and coauthors showed the physics of this knob in suspended graphene. They first cleaned devices to a mobility of about $1.2\times10^5$ cm²/V·s, then adsorbed short-range gases, which drove conductance at the charge-neutrality point from about $e^2/h$ to about $0.01\,e^2/h$ (Strong localization in suspended monolayer graphene by intervalley scattering). That study treated the disorder as an uncontrolled bulk variable. Gas adsorption cannot reach an h-BN-encapsulated channel, so we will introduce the dose through low-fluence irradiation. Dose will be recorded as fluence and treated as a nuisance-scaled variable rather than a known defect count.
On the experimental side, the devices follow the nano-ridge geometry of Topological Valley Transport in Bilayer Graphene Induced by Interlayer Sliding. The changes are at least four voltage probes at unequal spacings along the ridge, and nominally identical devices built with deliberately different contact metallization, so that contact systematics differ between devices. The full analysis, including priors, cuts and nuisance forms, is frozen on the first device and on simulation before data from the remaining devices are opened. This mirrors the unified blind processing used across separate runs in Dark Matter Search Results from 1.54 Tonne Year Exposure of PandaX-4T.
Prior work plays three different roles here. The sliding-channel and suspended-graphene papers motivate the project, since they show the gap. The neutrino papers supply the method, through baseline ratios and orthogonal channels. The first-principles transport studies constrain the interpretation. Calculations on line defects in MoS₂ (Effect of Line Defects on the Electrical Transport Properties of Monolayer MoS2 Sheet) and on twin boundaries in GaAs, where transmission is perfect at normal incidence but falls at finite transverse momentum (Effect of twin boundaries and structural polytypes on electron transport in GaAs), show that scattering strength depends on the incoming mode. A single $\lambda_{iv}$ is therefore an average over modes, and we will report it as an average over the eight channels rather than a per-channel quantity.
Baselines
The naive reference point is the attribution made implicitly in the sliding-channel work: every bit of excess resistance above $h/8e^2$ is intervalley scattering. That gives $\lambda_{iv} = L\,(h/8e^2)/\Delta R$ from a single probe pair. We will compute it from exactly the same raw data as our method. On the synthetic benchmark, where the truth is known, we expect it to be biased toward short $\lambda_{iv}$ whenever $R_s$ is nonzero, and to have no meaningful interval at all. Beating it means recovering the true $\lambda_{iv}$ with nominal 68% coverage where the naive estimate misses.
The strongest existing result on this question is the measurement in Topological Valley Transport in Bilayer Graphene Induced by Interlayer Sliding. It reports a ballistic mean free path of "approximately 1 μm" and conductance saturating near $h/8e^2$ up to about 80 K. We found no result on Lacuna that reports an intervalley mean free path with a stated uncertainty or a decomposition of the residual resistance. The concrete comparison is therefore whether our devices reproduce a mean free path of order 1 μm, and whether our analysis then splits that number into an intervalley part and a series or leakage part, each with an interval. Showing that a substantial fraction of the excess is length-independent would revise the earlier attribution, and that would count as beating the baseline.
The second strong competitor is a careful single-length fit with priors on contact resistance. This is what a diligent transport group would do without changing the device layout. We run it on the same data and give it the same nuisance priors and the same likelihood code. The only difference is that it sees a single probe spacing. This ablation isolates the length lever. Further ablations remove one orthogonal knob at a time, first displacement field, then temperature, then dose. Each is scored on the benchmark by bias, interval width and coverage, under both the parametric distortions and the unseen gap-closing distortion. A neural estimator in the FAIR Universe style, trained on the same simulations with the same budget, serves as a comparison for the inference engine, separate from the experimental design.
Why not just…?
Why not just fit weak-localization magnetoresistance, as graphene groups have done for years? Weak-localization formulas assume diffusive transport with small interference corrections. The valley channel is ballistic over about a micrometre, and the suspended-graphene devices sit deep in strong localization. In both regimes those formulas are invalid. Even where they apply, the phase-coherence and intervalley times trade off against each other in the fit. We will still run magnetoresistance on longer, diffusive control segments as a consistency check where the formula is valid.
Why not just use four-probe measurements, which already remove contact resistance? The sliding-channel device was already measured four-probe and still sat above the quantum limit. On a one-dimensional channel, voltage probes are themselves invasive, and four probes do nothing about ridge electrostatics or bulk leakage. Our single-length ablation is precisely the four-probe analysis with priors. If it matches the multi-length result on the benchmark, the length lever is unnecessary, and we will report that.
Why not just use the transfer-length method and stop? This alternative is a real threat, because the length slope is the backbone of our design. The difference lies in what the slope means. Gap-closing segments and bulk leakage also scale with length, so a bare slope can overstate intervalley scattering. The displacement-field, temperature and dose checks exist to show whether the slope is purely intervalley. The ablation that drops all three knobs measures how much they add. If that ablation already reaches nominal coverage under every simulated distortion, the honest conclusion is that the transfer-length method suffices for these devices, and the contribution shrinks to the symmetry argument plus the error budget.
Why not just apply neural simulation-based inference from the collider benchmarks? With about six nuisance parameters, the profile likelihood is cheap and transparent. The coverage problems that motivated FAIR Universe arise mostly from unmodeled shifts, and a better estimator does not cure those. The neural comparison on our benchmark will show whether it adds anything. We expect it to match the likelihood on parametric distortions and to share its failure on the unseen one.
Decisive Checks
The method stands or falls on four tests. First, on the synthetic benchmark, intervals must reach at least nominal 68% coverage across parametric distortions, and we report the coverage loss under the unseen gap-closing distortion. Second, in real devices, $R(L)$ must be linear in $L$ at fixed $D$ and $T$, and the fitted $\lambda_{iv}$ must stay constant as $D$ increases once the gap is open. A strong dependence on $D$ there would falsify the claim that the slope is intervalley-dominated. Third, under dose steps, $1/\lambda_{iv}$ must rise linearly while $R_s$ stays fixed within its interval. Fourth, after unblinding, devices with different contact metallizations must agree on $\lambda_{iv}$ at matched dose within their stated uncertainties, which is our counterpart of the cross-detector agreement Daya Bay used to certify its systematics. Agreement across all four would give intervalley scattering its first systematics-controlled measurement in a valley channel. Failure of the displacement-field or cross-device check would show that the "intervalley" excess in these devices is largely something else.
References
- Topological Valley Transport in Bilayer Graphene Induced by Interlayer Sliding
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- FAIR Universe – the challenge of handling uncertainties in fundamental science
- FAIR Universe HiggsML Uncertainty Dataset and Competition
- Strong localization in suspended monolayer graphene by intervalley scattering
- Dark Matter Search Results from <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mn>1.54</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>Tonne</mml:mi><mml:mo>·</mml:mo><mml:mi>Year</mml:mi></mml:mrow></mml:math> Exposure of PandaX-4T
- Effect of Line Defects on the Electrical Transport Properties of Monolayer MoS<inline-formula> <tex-math notation="LaTeX">$_{\bf 2}$</tex-math></inline-formula> Sheet
- Effect of twin boundaries and structural polytypes on electron transport in GaAs