When a neutrino hits an argon nucleus, the particles we eventually see in the detector are not the particles the neutrino made. On the way out, a pion can be swallowed by the nucleus, can turn into a neutral pion, or can knock out extra protons and neutrons, which often go unseen. Each of these detours carries energy away. Because experiments estimate the neutrino's energy from what survives, the estimate is biased, and oscillation physics depends on that estimate. This note proposes treating the nucleus the way indoor-positioning systems treat a building. The hidden history of each event is a walk through a small graph. Its nodes are hadron states, and its edges are the few physically allowed transitions: the "doors." Conservation laws play the role of walls that no path may cross. We calibrate the doors directly on hadron test-beam measurements on argon, then use the graph at analysis time to infer, event by event, which path the event most likely took and how much energy went missing along it.
The main claim is that this graph can be calibrated on data independent of neutrino interactions and then carried across target nuclei and across event generators. If so, it can correct energy estimates in ways that neither tuning a generator nor training a black-box regressor achieves. Indoor-localization work supplies the representational idea: noisy observations become interpretable once they are forced onto a graph of allowed transitions. Experimental-validation practice supplies the discipline: fit on one physical measurement, predict a different one, and report where the prediction breaks.
Core Claim
Current neutrino event generators treat final-state interactions, the rescattering of hadrons inside the nucleus, in one of two ways. The first is a full intranuclear cascade, a step-by-step Monte Carlo of hadrons moving through nuclear matter. The second is an "effective" model that assigns each hadron a single fate. The effective GENIE hA model was ruled out on argon at $p<0.001$ by ProtoDUNE-SP's exclusive π⁺–argon measurement. That measurement separated absorption from charge exchange: absorption was 160 ± 36 mb at 550 MeV, and charge exchange was 148 ± 52 mb. Tuned generators also fail elsewhere. MINERvA's multineutron measurement found MnvTunev1 about 50% too high at the peak, and no model reproduced the drop above $p_{T\mu}>0.6$ GeV/c.
We propose a third object, a transition graph. It is richer than a single fate, because it allows multi-step paths such as charge exchange followed by absorption. It is far smaller than a cascade, with tens of parameters. It is differentiable, and it is usable at inference time. Its nodes are hadronic states labeled by species, charge, and coarse kinetic-energy bin. Its edges are elastic escape, absorption, charge exchange, and nucleon knockout. Each edge carries a probability that depends on hadron kinetic energy $T$ and on nuclear size $A$.
The probability that a primary state $s$ becomes an observed final-state topology $f$ is
$$P(f \mid s;\theta) = \sum_{\pi:\, s \to f} \prod_{e \in \pi} p_e(T_e, A;\theta),$$
where $\pi$ ranges over allowed paths, $e$ over the edges on a path, $T_e$ is the hadron energy entering edge $e$, and $\theta$ are the edge parameters. The primary state is quasielastic, two-nucleon knockout (2p2h), or resonant pion production. The observed topology is, for example, "one muon, zero pions, one proton." The $A$-dependence enters only through a path-length factor scaling as $A^{1/3}$, the nuclear radius. This makes carbon-to-argon transfer a falsifiable prediction rather than a separate tune.
For a reconstructed event $x$, the graph gives a posterior over paths and hence an estimate of the energy that left no trace:
$$\hat E_\nu = E_{\text{vis}}(x) + \sum_{\pi} P(\pi \mid x;\theta)\, E_{\text{inv}}(\pi).$$
Here $E_{\text{vis}}$ is the calorimetric visible energy and $E_{\text{inv}}(\pi)$ is the binding energy, absorbed-pion energy, and neutron energy implied by path $\pi$. The quantity we care about is whether this correction reduces energy bias and resolution when the true physics differs from the physics used in training.
What the Indoor-Map Analogy Actually Contributes
Two indoor-space results carry over directly. Contact tracing over uncertain indoor positioning data showed a specific failure of distance in raw coordinates. People 1.5 m apart through a wall are far apart in walking distance, and ignoring walls produces false contacts. The nuclear version of this failure motivates our proposal. A zero-pion event from quasielastic scattering and a zero-pion event whose pion was absorbed can look alike in reconstructed kinematics. They are nevertheless separated by an absorption "door" that carries hundreds of MeV of invisible energy. A regressor that works purely on kinematic similarity treats them as neighbors.
The second result is GraphLoc. It matched a noisy graph built from WiFi signal jumps to a known floor-plan graph, reaching 91.3% node-mapping accuracy without a manual site survey. Its accuracy fell from above 87% in rooms to about 78% in open corridors, where boundaries blur. The nuclear analog of an open corridor is the low-energy regime. There, elastic and inelastic pion scattering overlap and the detector threshold hides soft protons. We therefore expect the graph prior to help most for events with clear exclusive topologies and least for low-recoil events, and we will report accuracy stratified this way rather than as one average.
Earlier work on sensor data in intelligent buildings makes the general point that physical boundaries serve as a prior that removes impossible trajectories. That is the role charge and baryon-number conservation play here.
Graph2Plan serves as a caution. Its layout graphs recorded which rooms were adjacent but not how people move between them, and it sometimes generated rooms with no entrance. We encode transitions, the doors, explicitly rather than mere co-occurrence of final-state particles.
Research Program
The first step is to build the graph and its calibration from hadron data alone. The ProtoDUNE-SP exclusive absorption, charge-exchange, and "other inelastic" cross sections at 550 and 750 MeV fix the argon pion edges in that energy range. We fit them by gradient descent through the path sum. This follows the analysis-by-synthesis approach of the differentiable LArTPC simulation, which fit entangled detector parameters to within 1% in closure tests but never touched nuclear physics or real data. Our graph is small enough that the 25-fold slowdown that paper reported for differentiable forward simulation is not a concern.
The ProtoDUNE data stop at 500–800 MeV, but neutrino-produced pions mostly lie below that range. The low-energy edges are therefore regularized with a smooth energy dependence and treated as the main uncertain extrapolation. We will quantify that uncertainty rather than hide it.
The second step is an identifiability analysis, which answers the experimental-validation concern about equifinality: different parameter sets can reproduce the same data. We compute the Jacobian of every published observable with respect to the parameters $\theta$ and the primary-channel normalizations. The observables are ProtoDUNE's exclusive channels, MicroBooNE's NC π⁰ 0p/1p split, the CC π⁰ rate, MINERvA's transverse kinematic imbalance, multineutron data, and A-dependence measurements. The rank and conditioning of this matrix tell us which combinations, such as 2p2h normalization versus pion absorption in zero-pion samples, are pinned by current data and which are not. The same matrix ranks candidate future measurements by how much degeneracy each would remove. This connects directly to a lever the neutrino community controls, namely which test-beam or exclusive-channel measurement to make next.
The third step is event-level inference. The graph takes reconstructed particles as input. These come from Pandora-style reconstruction (ProtoDUNE-SP Pandora) or from SparseSSNet labels. From them, the graph produces the path posterior and $\hat E_\nu$. The detector-level reconstruction is not our contribution. Hit-level graph networks such as Exa.TrkX address a different problem, linking detector hits into tracks. They are compatible with this work but are not a competitor.
Baselines
The naive reference point is calorimetric energy with a fixed per-topology offset for binding energy and missing hadrons. This is roughly what a standard analysis does before any model-dependent correction. Any method that fails to beat it on bias under generator mismatch is not worth its complexity.
The strongest forward-model competitor is a full cascade tuned to the same ProtoDUNE data through iterative reweighting. We will use the Geant4Reweight procedure from the ProtoDUNE-SP exclusive analysis, which reported reducing model-related bias to under 1% on the fitted channels. We will give it exactly the same input data, the same held-out observables, and an equal number of tuning iterations. Beating it means one of two things: better prediction on observables not used in fitting, especially the argon π⁰ 0p/1p split and the carbon-to-argon transfer, or equal prediction with a documented, identifiable parameter set. We will not claim to beat it on the fitted channels.
The strongest per-event competitor is a learned energy regressor. It is a network on reconstructed-particle features, trained on GENIE simulation, in the spirit of DUNE CVN. It gets identical inputs, the same training events, and a matched hyperparameter search budget. No reported energy-resolution result for such a regressor exists on Lacuna, so we will measure it ourselves.
The key ablation replaces the transition graph with a single-fate graph, with no multi-step paths and otherwise identical calibration. This isolates whether multi-step structure matters or whether calibrating hA-style fates on argon data would have sufficed. A second ablation removes the $A^{1/3}$ constraint and fits carbon and argon edges separately. This tests whether the shared structure carries real transfer information.
Why not just…?
Why not just retune the cascade with reweighting, as ProtoDUNE already did? This is the real threat for forward prediction, and we expect a well-tuned cascade to match our graph on the channels it was fitted to. Reweighting yields a better simulator, but it does not give a per-event, conservation-respecting posterior over what was lost. Its many coupled parameters also leave the 2p2h-versus-absorption degeneracy invisible rather than quantified. If the tuned cascade matches the graph on the held-out MicroBooNE split and the reweighted-cascade-plus-regressor matches our energy correction under generator mismatch, the graph's only remaining value is interpretability. We will say so.
Why not just train a neural network to regress neutrino energy from the event? Such a network absorbs the generator it was trained on. The DUNE CVN authors dropped their neutron-counting and neutrino/antineutrino outputs because the network could not handle these reliably. Our fake-data test settles the question: train on GENIE and evaluate on NuWro, NEUT, or GiBUU samples. When the physics shifts, the graph can be recalibrated on new hadron data by changing a few edge probabilities, while the regressor must be retrained on a simulation that may itself be wrong. If the regressor degrades no more than the graph under mismatch, the graph loses this argument.
Why not just calibrate hA-style single fates on argon data? The fact that hA was ruled out does not by itself show that fates are the wrong representation; it could reflect a poor tune. The single-fate ablation answers this directly. If single fates fitted to ProtoDUNE predict the MicroBooNE π⁰ split and transfer from carbon as well as multi-step paths do, the extra structure is unnecessary and we will adopt the simpler model.
Decisive Checks
We run three tests, each with a prediction that can fail.
The first is a held-out argon test. Calibrate on ProtoDUNE π⁺ data and carbon neutrino data, then predict MicroBooNE's NC π⁰ exclusive rates. These are (0.624 ± 0.131 syst ± 0.075 stat) × 10⁻³⁸ cm²/Ar for 0p and (0.444 ± 0.098 syst ± 0.047 stat) × 10⁻³⁸ cm²/Ar for 1p. Success means landing within the quoted uncertainties, where GENIE overpredicted the semi-inclusive rate of 1.243 × 10⁻³⁸ cm²/Ar.
The second is a transfer test. Fit on carbon (MINERvA's transverse imbalance and multineutron data), predict argon, then reverse. With the $A^{1/3}$ structure, the graph should beat the separate-fit ablation on the held-out nucleus.
The third is a fake-data energy test. Using GENIE, NEUT, NuWro, and GiBUU argon samples processed through the same reconstruction, compare energy bias and resolution for four methods under every train/test generator pair: the naive offset, the tuned cascade plus regressor, the learned regressor, and the graph. Results are stratified by final-state topology and by recoil energy, the "rooms versus corridors" split.
The graph should win on off-diagonal pairs, where train and test generators differ, and most clearly for exclusive topologies. If it wins only on diagonal pairs, or loses on low-recoil events by more than it gains elsewhere, the claim fails. In that case the identifiability analysis alone, a map of which nuclear-effect combinations current data cannot separate, becomes the deliverable.
References
- Measurement of Exclusive $π^+$--argon Interactions Using ProtoDUNE-SP
- Measurement of the multineutron <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mi>μ</mml:mi></mml:msub></mml:math> charged current differential cross section at low available energy on hydrocarbon
- Contact Tracing Over Uncertain Indoor Positioning Data
- GraphLoc: a graph-based method for indoor subarea localization with zero-configuration
- Graph2Plan: learning floorplan generation from layout graphs
- Differentiable simulation of a liquid argon time projection chamber
- Measurement of neutral current single <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> production on argon with the MicroBooNE detector
- First measurement of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>ν</mml:mi><mml:mi>μ</mml:mi></mml:msub></mml:math> charged-current <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> production on argon with the MicroBooNE detector
- Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions
- Measurement of the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>A</mml:mi></mml:math> dependence of the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>ν</mml:mi><mml:mi>μ</mml:mi></mml:msub></mml:math> charged-current quasielasticlike cross section as a function of muon and proton kinematics at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mo stretchy="false">⟨</mml:mo><mml:msub><mml:mrow><mml:mi>E</mml:mi></mml:mrow><mml:mrow><mml:mi>ν</mml:mi></mml:mrow></mml:msub><mml:mo stretchy="false">⟩</mml:mo><mml:mo>∼</mml:mo><mml:mn>6</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:mrow></mml:math>
- Reconstruction of interactions in the ProtoDUNE-SP detector with Pandora
- Semantic segmentation with a sparse convolutional neural network for event reconstruction in MicroBooNE
- Scaling Graph Neural Networks for Particle Track Reconstruction
- Neutrino interaction classification with a convolutional neural network in the DUNE far detector