Azimuthal Self-Calibration of Neutrino Detectors for Clean Nuclear-Effect Measurements

When a neutrino strikes an atomic nucleus, the physics of the collision is the same whichever way the event is turned around the beam axis. The detector that records the collision does not have this symmetry. Its wire planes, scintillator strips and drift field sit at fixed orientations, so a proton leaving the nucleus at one azimuth deposits charge differently from an identical proton at another. LiDAR mapping handles a similar problem: raw laser return strength mixes the surface's true reflectivity with range and viewing angle. Recent work shows that inverting the known physics of range and angle, then fitting a small data-driven correction on field data, gives a viewpoint-invariant quantity that downstream models use far better than raw intensity. We propose the same approach for neutrino detectors, with one addition. The rotational symmetry of the physics about the beam supplies a free "change of viewpoint" for every event. The detector's orientation-dependent response can therefore be calibrated on the neutrino data themselves, without trusting an event generator (the simulation program that models neutrino-nucleus collisions).

This matters because the observables that neutrino physicists use to study nuclear effects are measured through that same orientation-dependent response. The main example is transverse kinematic imbalance: the leftover momentum perpendicular to the beam after adding the outgoing muon and proton, which is nonzero only because the nucleus is not a free, stationary target. The clearest test case is an unexplained proton left-right asymmetry, $A_{LR} = -0.05 \pm 0.02$, reported in Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions. No generator reproduces it. The claim of this proposal is that an azimuth-invariance test can tell whether such an anomaly is nuclear physics or detector response, and that self-calibration against the symmetry gives cleaner measurements of binding energy and final-state effects.

Core Claim

A real nuclear effect must be independent of the event's azimuth $\phi$ about the beam axis, once the small breaking from beam divergence is modeled. A detector artifact generally is not, because $\phi$ determines how tracks are oriented relative to wires, strips and drift direction. The approach has two parts. First, we write the conversion from raw collected charge to energy loss per unit length ($dE/dx$) as an explicit physical inverse. This inverse corrects for drift-distance charge attenuation, which plays the role of LiDAR's range loss, and for the effective path length of a track across a wire or strip, which plays the role of the incidence-angle cosine. It also includes a recombination or quenching curve, the analog of LiDAR detector saturation. Second, we fit a low-dimensional residual correction $\eta$ over track orientation and drift position. The fit requires that standard-candle tracks with known energy loss, such as stopping muons and protons identified by range, give orientation-independent calibrated $dE/dx$. It also requires that the calibrated proton energy scale show no dependence on $\phi$. The correction is then frozen and applied to the nuclear-effect analysis.

The decisive new measurement is $A_{LR}$ and the transverse-imbalance peak location and width, each as a function of $\phi$. If these quantities are flat in $\phi$ and $A_{LR}$ survives, the evidence points to physics. If they are modulated in $\phi$, the evidence points to the detector.

Relation to Prior Work

The LiDAR side supplies the method. Reflectivity Is All You Need! inverts the $1/R^2$ and $\cos\alpha$ dependences and fits a near-range correction $\eta(R)$ from semantically labeled field data. Its semantic segmentation improves from 38.4% to 42.4% mean intersection-over-union, and it gains 6% in zero-shot transfer between sensors. We reuse its structure, a physical inverse plus a data-fitted residual, but replace semantic labels with a physical symmetry as the source of the constraint. A practical method utilizing multi-spectral LiDAR cancels geometry by taking ratios of co-registered channels. The analog in a multi-plane LArTPC (liquid argon time projection chamber) is to compare the same track across planes, which we use as an internal consistency check rather than the main method. CMD instead normalizes intensity statistically with a Box-Cox transform. We treat that as a baseline, not as the method.

The neutrino side supplies both the motivation and the closest competitor. Calorimetric classification of track-like signatures in liquid argon TPCs using MicroBooNE data already models $dE/dx$ as a function of residual range and local pitch, using likelihood densities built from simulation. It combines three wire planes for particle identification. This is the closest existing work. It differs from ours in three ways: its orientation correction is built from simulation rather than from a symmetry of the data, it is aimed at particle identification rather than energy scale, and it is never connected to nuclear-effect observables.

Three papers motivate the gap. The binding-energy study above shows that Fermi-gas models cannot simultaneously match the peak location, width and tails of the transverse imbalance, and it reports the $A_{LR}$ anomaly. Measurement of the $A$ dependence of the CCQE-like cross section finds generator breakdown growing with nuclear mass in exactly the transverse-imbalance regions we study. The multineutron cross-section measurement finds the tuned model over-predicting multi-neutron production by about 50%, which shows how far generators can be from data.

One paper constrains the proposal. Measurement of the axial vector form factor from antineutrino–proton scattering isolates scattering on free hydrogen in hydrocarbon using kinematic constraints. We borrow that selection as a null control, since free-proton events carry no nuclear imbalance. We also accept its limitation: argon contains no free hydrogen, so in argon only the symmetry constraint is available.

Research Program

The first step is a simulation study that measures how much detector response can fake nuclear effects. Using Geant4-level detector simulation of a MINERvA-like scintillator and a MicroBooNE-like three-plane LArTPC, we inject controlled orientation-dependent response distortions of plausible size. We then measure how much spurious $A_{LR}$, peak shift and peak broadening each distortion produces in the transverse imbalance. This tells us whether the observed $-0.05$ is even reachable by detector effects. If it is not, the proposal's main diagnostic question closes early, and the work reduces to calibration for precision.

The second step builds the calibration. The physical inverse has a handful of parameters: attenuation length, recombination or Birks quenching constants, and the path-length geometry. The residual $\eta$ is a smooth, low-order expansion in track orientation relative to each readout view and in drift position. The fit loss combines three terms: standard-candle orientation invariance, cross-plane agreement for the same track, and flatness of the calibrated proton energy scale in $\phi$.

This last term carries a subtlety that could break the approach. Detector acceptance also varies with $\phi$, because tracks running parallel to wires are lost more often. The calibration must therefore constrain per-track energy response, never event-count distributions. Otherwise acceptance differences leak into the energy scale.

The third step applies the frozen calibration to the nuclear observables. These are $A_{LR}(\phi)$ and the $\delta p_{Ty}$ peak location and width, which carry the binding-energy information, measured on carbon-like, argon-like and heavier targets. All real-data steps require collaboration-level access to reconstructed events. Until that access exists, the program runs on simulation with injected effects, and this dependency is stated up front.

Baselines

The naive reference point is the standard calibration used in the published analyses, applied unchanged. Every metric below is reported relative to it.

The strongest existing method is the pitch-dependent, simulation-derived likelihood approach of the MicroBooNE calorimetric paper. On its own task it reports 93.7% proton efficiency at 10% muon mis-identification and an area under the ROC curve of 0.977, compared with 0.955 for the collection plane alone. We will reproduce its orientation-dependent correction on the same simulated samples and use it as the response model feeding the kinematic analysis. Both methods get the same standard-candle samples and the same number of free residual parameters. To beat it, our calibration must reach three things: lower orientation-dependent proton energy bias against simulation truth, smaller $\phi$-modulation of the energy scale when the simulated response is deliberately mis-modeled relative to the "data" sample, and no loss in particle-identification AUC below 0.977. The mis-modeled case is where the simulation-derived method should fail and the symmetry method should not.

A statistical baseline applies Box-Cox or quantile normalization of charge per orientation bin, following CMD. In LiDAR detection that style of approach reached a 42.89% mean closed gap, against 9.23% for self-training. We run it on identical bins and samples.

A learned baseline is a 3D convolutional regressor in the style of Deep-Learning-Based Kinematic Reconstruction for DUNE, which reported a 31% RMS improvement in neutrino energy reconstruction. We train it with equal compute on raw inputs and on calibrated inputs. The comparison is made under a generator swap, training on one generator and testing on another, and under a detector-response swap.

Two ablations isolate the key ingredient. The first removes the $\phi$-flatness term and keeps only the physical inverse and standard candles. The second removes the physical inverse and fits $\eta$ alone.

Why not just…?

Why not just normalize the charge distribution in each angle bin statistically? This is the most serious alternative, because the physics really is $\phi$-invariant, so distribution matching is not obviously wrong. The problem is that it cannot separate acceptance from response. When long-along-wire tracks drop out of the sample, the energy spectrum in those $\phi$ bins changes, and quantile matching absorbs that change into a false energy-scale shift. It also carries no per-track meaning that transfers across particle species. The simulation experiment will decide the question: with injected acceptance holes and response distortions of known size, we measure whether the statistical baseline recovers the true proton energy. If it matches our method, we will adopt it and say so.

Why not just let an end-to-end network learn the orientation dependence? Networks trained on simulation learn the simulation's detector model. Liu et al. note energy-dependent biases that required reweighting the training sample. In the LiDAR case, feeding raw intensity to strong networks gave 38.4% mean intersection-over-union against 42.4% with calibrated reflectivity. The response-swap test measures directly whether the network's reconstructed transverse imbalance picks up a spurious $A_{LR}$ when the real detector differs from the simulated one.

Why not just tune the generator to the transverse-imbalance data? If part of $A_{LR}$ or the peak shift comes from the detector, tuning nuclear parameters to fit it builds a detector artifact into the nuclear model. That error would then propagate to argon oscillation analyses. The binding-energy study already shows that tuning cannot fit peak location and width together, and the $\phi$-resolved measurement tells us which residual is legitimately nuclear before any tuning is done.

Why not just use free-hydrogen events? They are an excellent null test, and we use them as one. They cannot serve as the calibration in argon detectors, which contain no free protons.

Decisive Checks

Four checks decide the outcome. First, calibration must preserve genuine physics: when we inject a known physical left-right asymmetry into simulation by reweighting, it must survive calibration within statistical error. Second, when we inject a detector-only asymmetry, calibration must remove it. Third, on hydrocarbon data or simulation, the free-hydrogen sample selected with the axial-form-factor kinematics must give $A_{LR}$ consistent with zero after calibration; a nonzero value there is purely detector-driven. Fourth, the Fourier amplitudes of $A_{LR}(\phi)$ and of the $\delta p_{Ty}$ peak position must be consistent with zero after calibration and with beam-divergence modeling included.

If $A_{LR}$ remains at about $-0.05$ with a flat $\phi$ dependence and a null hydrogen control, it becomes a much stronger case for new nuclear physics. If it tracks detector orientation, a long-standing anomaly is explained and removed from the list of targets for generator tuning.

References

  1. Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions
  2. Reflectivity Is All You Need!: Advancing LiDAR Semantic Segmentation
  3. A practical method utilizing multi-spectral LiDAR to aid points cloud matching in SLAM
  4. CMD: A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection
  5. Calorimetric classification of track-like signatures in liquid argon TPCs using MicroBooNE data
  6. 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>
  7. 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
  8. Measurement of the axial vector form factor from antineutrino–proton scattering
  9. Deep-Learning-Based Kinematic Reconstruction for DUNE