Mixture-Anchored Differential Corrections for Neutrino–Nucleus Event Generators

Neutrino experiments turn detector hits into neutrino energies using simulations of how a neutrino strikes a nucleus. Those simulations are known to be wrong in a specific way. Protons and neutrons inside a nucleus sometimes form tightly bound, fast-moving pairs, called short-range correlations. These pairs hold roughly a fifth of the nuclear strength, and the simulations under-describe them. Retuning a simulation's knobs cannot fix this, because the fault lies in the model's structure rather than its settings. One wants to correct the simulation directly to data, but the data that isolate these effects are few and noisy. A flexible correction then ends up fitting noise.

Speech engineering met the same shape of problem in voice conversion, the task of making one speaker sound like another. It learned two lessons. First, divide the input space into a few soft regions with a Gaussian mixture (a sum of bell-shaped clusters), and learn a simple local shift in each region. Second, apply that learned difference to the original signal instead of rebuilding the signal from a coarse parametric model. We propose to carry both lessons to neutrino physics. The correction moves and reweights simulated events inside a small number of physically anchored regions: the quasi-elastic peak, the two-nucleon tail, and the region dominated by rescattering inside the nucleus. It is fit to measured distributions of transverse kinematic imbalance, and it is sized so that its parameters can actually be resolved at current exposure.

Core Claim

A correction written as a soft mixture of local affine transports, fit to published differential cross sections with their covariances, can do three things. It can capture the data–simulation gap that generator retuning cannot reach. It can transfer that fix to kinematic variables it was never fit on. And it can do so with few enough parameters that its uncertainty is limited by counting statistics rather than by overfitting.

The neutrino-energy bias it predicts should be more accurate than that of a tuned generator, and more stable than that of flexible reweighting methods. The main claim is testable with generator-swap mock data. The size of the correction on real data is checked directly against MINERvA's two-variable measurements.

Why This Gap Is Real

Three kinds of evidence motivate the work.

The first comes from electron scattering. Knockout experiments at Jefferson Lab Hall A on argon and titanium find fewer protons than the independent-particle picture predicts. About 20% of the strength sits in correlated pairs (Determination of the titanium spectral function from (e,e'p) data; Determination of the argon spectral function from (e,e'p) data).

The second comes from neutrino scattering. MINERvA built imbalance variables that separate Fermi motion, the internal motion of nucleons, from multi-nucleon effects. Its results show three failures:

The third concerns the reconstruction networks built downstream. The DUNE and NOvA networks inherit whatever their simulation assumes (Deep-Learning-Based Kinematic Reconstruction for DUNE; Electron Neutrino Energy Reconstruction in NOvA Using CNN Particle IDs).

The same MINERvA papers also constrain the design. Both state that statistical uncertainty in fine kinematic slices limits how well competing models can be told apart. Any correction with many free parameters will therefore fit Poisson fluctuations. The unfolding literature makes the same warning about under-regularized inversions producing false structure (Unfolding Detector Distortions in Physical Spectra).

The research file flags a structural link between mixture-based conditional regression in voice conversion and nuclear spectral functions (Statistical Voice Conversion using Gaussian Mixture Models). It also records that this link has not been used. The obvious version, a joint-density model trained on paired source and target frames, does not transfer: simulated and real neutrino events are not paired. We therefore take the non-parallel, difference-based form of the idea. The closest precedent is the NU-NAIST system, which filtered the source waveform with learned spectral differences instead of re-synthesizing through a vocoder (a parametric speech synthesizer). It reached about 74% speaker-similarity accuracy with a mean opinion score above 3 in the Voice Conversion Challenge 2016 (The NU-NAIST Voice Conversion System for the Voice Conversion Challenge 2016; The Voice Conversion Challenge 2016). It supplies the method template. It says nothing about physics.

Research Program

The correction model

We fit a Gaussian mixture with $M$ components to the generator's true final-state variables for mesonless muon-plus-proton events. Write these variables as $x$: muon momentum and angle, proton momentum and angle, and the derived quantities $\delta p_T$ and $\delta\alpha_T$.

Each simulated event is moved and reweighted:

$$x' = x + \sum_{m=1}^{M} P(m \mid x)\left[A_m (x - \mu_m) + b_m\right], \qquad w' = w \sum_{m=1}^{M} P(m \mid x)\, e^{c_m}.$$

The symbols are as follows:

  • $P(m \mid x)$ is the probability that the event belongs to component $m$.
  • $\mu_m$ is the center of component $m$.
  • $A_m$ is a diagonal stretch matrix and $b_m$ is a shift vector for component $m$.
  • $c_m$ is a log-normalization for component $m$. It carries the missing strength that pure transport cannot supply.
  • $w$ is the event's original weight.

This mirrors differential voice conversion in one key respect. Only the difference is learned. Every other property of each event is kept unchanged, including its correlations with unfitted variables such as hadronic energy and neutron content.

Fitting

The parameters $\theta = \{A_m, b_m, c_m\}$ are fit by minimizing the covariance-weighted $\chi^2$. This compares the corrected generator's histograms with MINERvA's published unfolded differential cross sections in $\delta p_T$ and $\delta\alpha_T$, using the full covariance. Generator events are produced and binned through NUISANCE, following the observable-topology convention of Comparisons and challenges of modern neutrino scattering experiments (TENSIONS2016 report).

The number of components $M$ is chosen by held-out $\chi^2$ under bootstrap resampling of the data covariance, not by training fit. This guards against fitting fluctuations.

Resolvability analysis

We compute the Fisher information of each parameter. We then scale the statistical part of the covariance as 1/POT, where POT (protons on target) measures beam exposure. The 1/POT scaling holds because the counts are Poisson. From this we report the exposure at which each component's correction becomes distinguishable from zero. This makes explicit which parts of the multi-nucleon correction current data can actually support.

Over-smoothing

Voice conversion showed that conditional-mean mappings squeeze variance and wash out sharp peaks (Alleviating the over-smoothing problem in GMM-based voice conversion with discriminative training). The affine form above lets $A_m$ widen or narrow each region independently. As an ablation, we compare it against a shift-only version ($A_m = I$), which is prone to smearing the quasi-elastic peak width.

Baselines

Uncorrected generators. The naive reference is the stock GENIE, NuWro and GiBUU configurations, compared to the same MINERvA data through NUISANCE. For goodness of fit on MINERvA's $\delta p_T$ and $\delta\alpha_T$ data, we found no reported result on Lacuna. The only related number is qualitative: the imbalance paper reports systematic underprediction at intermediate momentum transfer. We will compute the stock $\chi^2$ values ourselves, under the identical covariance.

Tuned generator. The strongest existing competitor is a tuned generator: MnvGENIE-v1 with its two-nucleon enhancement, plus the binding-energy shift that Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions found improves GENIE. This is what experiments actually deploy, and it has physical interpretability. For fairness, we will re-tune its knobs on exactly the same $\delta p_T$ and $\delta\alpha_T$ training data, with a parameter budget at least matching ours.

Flexible reweighting. A regularized classifier-based reweighter, a neural network that learns per-event weights on the same true variables, is trained to the same data. Its regularization strength is chosen by the same held-out procedure and given the same tuning budget.

Forward folding. The spectral-function forward-folding approach of Determination of the argon spectral function from (e,e'p) data serves as a reference for parametric physics fits. It reported $\chi^2/\text{d.o.f.} = 1.9$ on argon reduced cross sections. That is a different observable, so it is a benchmark for fit quality in this literature rather than a head-to-head competitor.

Ablations. Two ablations isolate the key ingredients. A reweight-only variant ($A_m = I$, $b_m = 0$) tests whether moving events matters or whether normalization alone suffices. A single-component variant ($M = 1$) tests whether the physical anchoring by region matters.

What counts as a win. Beating the tuned generator means two things on generator-swap mock data. First, a lower held-out $\chi^2$ on variables that were not fit. Second, a smaller bias in the reconstructed-minus-true neutrino-energy distribution. Beating the reweighter means matching its held-out accuracy with smaller bootstrap variance of its predictions.

Why not just…?

Why not just retune the generator? Retuning is limited to the model's own parameter space. The file's evidence says that space does not contain the data. Fermi-gas binding knobs cannot fix the peak, the width and the tail at once. Even tuned MnvGENIE-v1 underpredicts intermediate momentum transfer. If a retune nonetheless reaches our held-out $\chi^2$ on real data and in generator-swap closure, then the differential correction adds nothing and we will say so. The first experiment is designed to settle exactly this.

Why not just reweight events with a flexible classifier? Reweighting cannot move events. A peak shifted by binding energy is a displacement, and reweighting can only mimic it by suppressing one side and boosting the other. That inflates the effective variance. Flexible reweighters also have many degrees of freedom against Poisson-limited data. This is a real threat when the regularization is good. The bootstrap-variance comparison and the held-out-variable transfer test will show whether the mixture's economy buys stability.

Why not just unfold the data and compare models in true space? Unfolding corrects the data histograms, not the simulation. It therefore gives no event-level correction that can propagate to neutrino energy or retrain a network. It also brings in its own prior dependence. Our method consumes the unfolded results that MINERvA already publishes, so the two approaches complement each other. The prior-dependence concern is addressed by also running the fit on forward-folded mock data.

Proposed Experiments

1. Generator-swap closure

This is the decisive check. We treat NuWro with a spectral function, and separately GiBUU, as "truth." We generate mock MINERvA-exposure data at 3.28×10^20 POT with realistic covariance. We then fit all methods starting from GENIE, using only $\delta p_T$ and $\delta\alpha_T$.

Evaluation is on quantities the fit never saw:

  • the $\delta p_{Tx}$ and $\delta p_{Ty}$ projections;
  • the proton left-right asymmetry, measured in real data as $A_{LR} = -0.05 \pm 0.02$;
  • the true neutrino-energy residual distribution.

If our correction does not beat the tuned generator here, the proposal fails.

2. Real-data transfer

We fit on MINERvA's measured $\delta p_T$ and $\delta\alpha_T$ and predict the separately published $\delta p_{Tx}$ and $\delta p_{Ty}$ distributions and $A_{LR}$. We score these predictions with the published covariances. The $A_{LR}$ asymmetry is unmodeled by generators, so a correct sign and magnitude would be informative.

3. Resolvability scan

We report, per mixture component, the exposure at which its correction exceeds twice its uncertainty. We expect the quasi-elastic-peak shift to be resolved now. We expect the intermediate-momentum-transfer tail to be marginal, consistent with the MINERvA authors' own statement that statistics limit model discrimination there.

4. Downstream stress test

We retrain a simple energy regressor on corrected versus uncorrected GENIE and evaluate it on NuWro mock data. This measures how much of the bias exported through neural estimators is removed. The simulated DUNE benchmark reports a 31% energy-resolution gain for these estimators (Deep-Learning-Based Kinematic Reconstruction for DUNE), but that gain is measured only against the same simulation, and this test checks the other side.

References

  1. Determination of the titanium spectral function from <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:mi>e</mml:mi><mml:mo>,</mml:mo><mml:mtext> </mml:mtext><mml:msup><mml:mrow><mml:mi>e</mml:mi></mml:mrow><mml:mrow><mml:mo>′</mml:mo></mml:mrow></mml:msup><mml:mi>p</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:mrow></mml:math> data
  2. Determination of the argon spectral function from <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mo stretchy="false">(</mml:mo><mml:mi>e</mml:mi><mml:mo>,</mml:mo><mml:msup><mml:mi>e</mml:mi><mml:mo>′</mml:mo></mml:msup><mml:mi>p</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math> data
  3. Measurement of Final-State Correlations in Neutrino Muon-Proton Mesonless Production on Hydrocarbon 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>3</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:mrow></mml:math>
  4. Nucleon binding energy and transverse momentum imbalance in neutrino-nucleus reactions
  5. 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
  6. Deep-Learning-Based Kinematic Reconstruction for DUNE
  7. Electron Neutrino Energy Reconstruction in NOvA Using CNN Particle IDs
  8. The NU-NAIST Voice Conversion System for the Voice Conversion Challenge 2016
  9. The Voice Conversion Challenge 2016
  10. Comparisons and challenges of modern neutrino scattering experiments (TENSIONS2016 report)
  11. Alleviating the over-smoothing problem in GMM-based voice conversion with discriminative training