WorksPhysics and Astronomy
Machine-learning-based neutrino-event generation constrained by experimental measurements
Abstract
We present a generative surrogate for charged-current -nucleus scattering that can be calibrated to measured differential cross sections without event-level experimental samples. Starting from a theoretical interaction model, we fine-tune the surrogate to a MicroBooNE measurement and show that this also improves the description of observables not used in the fit. The calibrated model can also be applied to measurements from different neutrino fluxes without retuning, while retaining event-level predictions of outgoing-muon kinematics. Our results show that experimental cross sections can directly constrain generative event models while preserving their applicability across different observables and beam configurations.
Introduction. Precision measurements of neutrino oscillations and leptonic CP violation are demanding tests for establishing the Standard Model’s consistency. They may also shed light on the matter-antimatter asymmetry of the Universe [1]. To achieve these precision measurements at next-generation long-baseline neutrino experiments, such as DUNE [2] and Hyper-Kamiokande [3], an accurate modeling of neutrino–nucleus interactions is required. Nuclear effects matter because they affect final states and neutrino-energy reconstruction [4, 5]. Recent studies show that uncertainties in final-state interaction modeling can induce distortions comparable to the precision required for oscillation-parameter measurements [6]. At the same time, increasingly sophisticated theoretical neutrino interaction models make efficient event generation in Monte Carlo (MC) simulation tools increasingly challenging [11, 12, 13] [7, 8, 9, 10]. Generative deep-learning models provide natural surrogates for stochastic particle–nucleus scattering, learning multidimensional distributions of final-state observables and allowing efficient event sampling [14, 15, 16, 17, 18]. Beyond reproducing a single process, transfer learning can extend such models to related processes [19].
Accelerator neutrino beams are not monoenergetic but have a broad, energy-dependent flux [20]. Measured cross sections are flux-averaged, with several reaction mechanisms contributing simultaneously, complicating both model validation and MC event-generator tuning. For a fixed nuclear target, however, interaction physics, nuclear effects, and cross section distributions depend on , independently of how the neutrinos were produced. Detector response, acceptance, and backgrounds are experiment-specific and enter the measurement and its uncertainties. At the interaction level, however, experiments using the same nuclear target differ primarily through their neutrino flux . A surrogate conditioned on can therefore transfer interaction information between beam configurations, with the beam profile as the experiment-specific input.
We realize this with a generative adversarial network (GAN) which readily learns unnormalized event distributions for neutrino and antineutrino scattering off nuclei [17, 18], but an unnormalized shape cannot be compared to a measured cross section nor transported to a different beam. The missing ingredient is the energy-dependent interaction probability. This weighting is precisely what a bare generator lacks. We therefore couple the GAN to a deep neural network (DNN) that predicts the total cross section , so we can normalize generated events to the cross section. The surrogate can then be fine-tuned directly on flux-averaged measurements, and its interaction-level predictions can be transported to other experiments by supplying the corresponding neutrino flux, enabling a direct comparison between the surrogate and measurements not used in the training. In this sense, the framework provides a way to fine-tune an MC interaction model directly from the measured cross sections, without requiring event-level experimental information. This differs from machine-learning unfolding approaches such as OmniFold [21], which iteratively reweight simulated events using detector-level data and detector simulation. Instead, our approach modifies the generative model and is constrained directly by binned cross sections and their covariance, rather than per-event detector-level samples. It also differs from the parametric generator tunes [22], which adjust a predefined set of parameters of the underlying interaction model.
The central idea is to retain the event-level structure learned from a physics-based MC model while allowing measured cross sections to modify its predictions. We therefore couple a -DNN, which supplies the energy-dependent normalization, to a conditional GAN that generates the outgoing-muon kinematics at fixed . Relative to Ref. [17], the key extensions are the explicit normalization for arbitrary neutrino fluxes and direct fine-tuning to measured cross sections rather than simulated events alone. The resulting surrogate therefore combines MC-based event structure with direct experimental constraints and can be applied across different beam configurations.
We test this strategy by calibrating the model to a MicroBooNE measurement [23] of the inclusive charged-current (CC) -argon scattering double-differential cross section , and evaluating it against measurements involving different observables and neutrino fluxes. The tuned surrogate improves the description of MicroBooNE observables not used in the training [24, 25, 26] and is consistent with the ArgoNeuT measurements [27, 28] obtained with a substantially different flux. These results show that experimentally measured cross sections can constrain a generative event model while remaining applicable across different observables and beam configurations.
We construct our model in two stages. First, we obtain the -DNN and GAN models independently from data produced with the NuWro MC event generator. Then we fine-tune both models simultaneously using experimental data. Figure 1 illustrates the model optimization scheme.
Figure 1. Training scheme: (a) The -DNN and GAN are pre-trained on NuWro. (b) Fine-tuning alternates adversarial regularization using the data-informed reservoir with a differentiable full-covariance fit to the forward-smeared MicroBooNE measurement. The latter updates and the -DNN.
The first stage. The -DNN is a feed-forward fully connected network. Rather than mapping a single energy to a cross section, it predicts the cross section averaged over an energy interval, with the single-energy value recovered for a vanishing interval width.
Using the -DNN and a known experimental flux (in our case, from MicroBooNE), we obtain a sample of interacting neutrino energies. Namely, the total-cross-section-weighted energy distribution of incoming neutrino events reads: , where is the total cross section predicted by the -DNN (or by NuWro) (averaged over the -th bin), and , is the flux-averaged cross section, where the flux satisfies: , and is the domain covered by the flux.
To generate events, we use a GAN conditioned on the neutrino energy to reproduce the muon kinematics as a function of . The generator maps a latent noise vector, together with , to the muon energy and scattering angle ; the discriminator learns to distinguish these generated events from NuWro training events. We trained the network from scratch on CC -argon events generated with NuWro in the same manner as in Ref. [18]. Technical details of the -DNN, GAN architecture and training, fine-tuning procedure, differentiable histogramming, and additional comparisons are given in the Supplemental Material (SM) [29].
Using the event distribution , we randomly sample neutrino energies for the GAN generator to obtain cross-section-weighted event samples. Differential cross sections are then reconstructed from the generated event fractions and the corresponding -DNN normalization; the explicit construction used in the fine-tuning is given in the SM.
In the second stage, we fine-tune the -DNN and GAN to the MicroBooNE CC -argon measurement [23]. We achieve this by fine-tuning the models using a nonstandard procedure. The difficulty
is twofold: the MicroBooNE data are presented as a three-dimensional differential cross-section distribution (double-differential in bins of ), whereas the GAN is trained on event samples; and the differential distributions also encode the total cross section, which in our system is reproduced by the -DNN. Both observations lead us to a joint optimization procedure in which the GAN and -DNN weights are updated together using the experimental distributions.
The second-stage optimization combines two complementary sub-steps. In the first, the discriminator is fine-tuned on events drawn from a reservoir, while the generator is updated adversarially against the discriminator; only the final layers of both networks are allowed to adapt. The generator samples neutrino energies from the flux reweighted by the current -DNN prediction.
The reservoir consists of one million truth-level NuWro events resampled from a larger sample generated with the MicroBooNE flux. The MicroBooNE result is a smeared double-differential cross section in , reported in bins of , with the smearing matrix relating the truth-level prediction to the measured bin space. The measured bin contents, together with the MicroBooNE flux, determine non-negative target occupancies for the reservoir, as detailed in the SM. Since the reservoir is constructed from truth-level NuWro events using a smeared measurement, it cannot be interpreted as a truth-level target; rather, it regularizes the generator toward the physical, NuWro-like manifold. We perform the quantitative fit separately in the second sub-step, where we forward-smear the generated prediction with and compare it with the original measurement.
In the second sub-step, we update the generator and the -DNN jointly using the differential cross-section measurement. We histogram events generated by , and normalize each histogram by the -DNN prediction for the flux-averaged cross section over the corresponding neutrino-energy interval . This yields an estimate of the double-differential cross section in each interval. The prediction is then forward-smeared, , and compared with the measured through a constructed with the full covariance matrix; see Eq. S-IV.13 of the SM.
The full covariance is retained because the unfolding and the smearing induce bin-to-bin correlations. If we do not follow this prescription, the fitting procedure collapses, as discussed in Ref. [30]. To make the comparison differentiable with respect to the network weights, the histogramming step is replaced by a smooth approximation, detailed in Sec. S-IV B of the SM. The sampling of conditions the generation but does not contribute to the gradient; the normalization through does.
| -DNN, GAN | 461.7/138 | 460.3/138 |
| -DNN, GAN | 435.6/138 | 268.4/138 |
| -DNN, GAN | 152.0/138 | 500.1/138 |
| -DNN, GAN | 110.5/138 | 220.3/138 |
Table I. The values of and over the number of bins (n) for four configurations of the models: base (subscript B) and fine-tuned (subscript T) -DNN and GAN models. Computed for samples of 1 million events.
Table I summarizes the effectiveness of the fine-tuning, while Figure 2 illustrates the resulting agreement with the data. We report two complementary metrics: the diagonal , defined in Eq. (S-IV.14) of the SM, which primarily probes per-bin agreement, and the full-covariance , which also accounts for the bin-to-bin correlations induced by the unfolding and the smearing matrix . The table shows these metrics for the four possible post-training combinations of the base and tuned -DNN and GAN components.
Figure 2. Base and tuned models compared with the MicroBooNE double-differential cross-section measurement [23] for various ranges. Grey vertical dashed lines separate different regions, from left to right: [0.2, 0.705], [0.705, 1.05], [1.05, 1.57], and [1.57, 4] GeV.
Replacing only the base -DNN with its tuned counterpart lowers sharply but does not improve , indicating that this component mainly corrects the cross-section normalization. Conversely, replacing only the base GAN with the tuned generator substantially improves while leaving nearly unchanged, showing its dominant role in reproducing correlated kinematic structure. Only the combination of both tuned components improves both
metrics simultaneously, demonstrating their complementary roles and motivating their joint optimization.
We note that the first sub-step, in which the discriminator is exposed to the reservoir, is essential. Without it, training with the loss term alone reproduces the smeared, flux-averaged distributions but distorts the quasielastic (QE) and -resonance peaks; see Figure 3: the smearing and the coarse measurement binning wash out this sharp structure in the observable to which is fit, so the covariance objective alone cannot constrain it. The reservoir, built from truth-level NuWro events that carry this structure, supplies exactly the differential information the smeared lacks, keeping the generator on the physical manifold of reaction events.
Figure 3. Event distribution at GeV generated by the fine-tuned GAN with and without the reservoir sub-step.
We next test whether the tuned surrogate is consistent with other published MicroBooNE cross-section measurements not used in the fit. We first consider an earlier MicroBooNE measurement [24] of the total cross section and the differential cross sections and , with . The model was tuned only on the cross section of [23]; the observables considered here are therefore an out-of-training test of what the surrogate has learned, not a refit. Figure 4 shows the corresponding cross sections. The tuned model improves the description of all three observables relative to the base model (both in the shape and in the normalization): in the full-covariance metric improves from 58.1/10 to 44.2/10 for , from 29.1/11 to 12.7/11 for , and from 17.1/8 to 4.4/8 for . The improvement is most pronounced at neutrino energies below 2 GeV and muon energies below 1 GeV, the region best constrained by the data.
Figure 4. Base and fine-tuned models compared to the MicroBooNE measurement [24]: (left), (center) and (right).
Finally, we compare our model against MicroBooNE and ArgoNeuT data not used in training. We start from more recent MicroBooNE measurements [25, 26] of cross section shown in Figure 5. The comparison for the double-differential cross section [25, 26] is shown in SM. These measurements were not used at any stage of training or model selection; they therefore constitute a cross-measurement test of the surrogate. The tuned model improves markedly over the base model in both observables. In the full-covariance metric, drops from 257.0/69 to 109.2/69 for , and from 291.5/17 to 6.8/17 for . The latter is a notable result: a model tuned on a single, distinct measurement predicts a later measurement not used in training at , without any retuning. This shows that the data-driven adaptation
Figure 5. Base model, tuned model, and the MicroBooNE data [25, 26] for , not used in the training.
remains consistent with a different representation of the same underlying interaction kinematics.
ArgoNeuT provides a qualitatively different test: it uses the NuMI beam in its low-energy configuration [31], whose flux extends well above the energies covered by the training data, and it is a distinct detector. We therefore use it primarily to probe the extrapolation of the framework rather than the data-driven adaptation within the constrained region. We obtain predictions with NuWro and our DNN using the fluxes of Refs. [27, 28]. ArgoNeuT did not publish the covariance matrix, so a full cannot be formed; we report the diagonal only as an indicator, noting that it neglects the bin-to-bin correlations of the measurement.
Figure 6 shows the distribution for ArgoNeuT; see also the SM for the antineutrino-mode results. Because the ArgoNeuT flux lies almost entirely above the energy range constrained by the MicroBooNE fine-tuning, the tuned and base models give similar flux-averaged predictions, and the agreement with the data is therefore inherited mainly from the base model’s high-energy behavior. Both models yield , although this partly reflects the large ArgoNeuT uncertainties. Within these limitations, the tuned surrogate remains consistent with the out-of-training ArgoNeuT data, indicating that the fine-tuning does not substantially degrade the high-energy behavior inherited from the base model.
Figure 6. Base model, tuned model, and ArgoNeuT data [27, 28] for distribution. The shaded area shows the range of the training MicroBooNE data.
In summary, we have constructed a machine-learning-based neutrino-event generator that can be calibrated directly to binned experimental cross sections without event-level experimental samples. A conditional GAN generates the outgoing-muon kinematics (energy and scattering angle ), while a -DNN provides the energy-dependent absolute normalization; both are jointly constrained by the forward-smeared MicroBooNE measurement. The central result is that experimental data can update a generator pretrained on a conventional Monte Carlo model, and it can then be used for different observables and flux configurations.
A model tuned to the MicroBooNE measurement improves the description of other MicroBooNE observables not used in training and remains consistent with ArgoNeuT data obtained with a substantially different flux. The data-constrained region modifies the normalization and correlated kinematics, while outside this region the prediction largely inherits the underlying MC model. The framework therefore defines a route toward event generators that combine theory-based simulations with direct experimental constraints; extending the constrained kinematic range and propagating calibrated model uncertainties are natural next steps. With the latter in place, the surrogate could serve as a common interaction reference that, folded with each beam’s flux, tests the mutual consistency of cross-section measurements across experiments.
The fine-tuned models and the event-generation script are available in Ref. [32].
Acknowledgments
This research is partly (K.M.G., B.E.K., J.T.S.) or fully (J.L.B., H.P., R.D.B.) supported by the National Science Centre under grant UMO-2021/41/B/ST2/02778. J.L.B. and J.T.S. are also supported by the Polish Ministry of Science Grant No. 2022/WK/15.
Model optimization
In the following sections, we describe in detail all steps of the model optimization. Figure S1 presents an expanded version of the model-optimization scheme shown in Figure 1 of the main text.
Figure S1. Training scheme. (a) The -DNN and the conditional GAN (, ) are first trained on NuWro. (b) Fine-tuning combines (1) adversarial regularization using the data-informed reservoir and (2) a differentiable full-covariance fit in the measurement space. The discriminator is updated on every reservoir batch, while every fifth discriminator update is followed by an adversarial update of and by the update of and the -DNN. Sampling of from carries no gradient.
Cross-section DNN: parametrization and training
The -DNN is a feed-forward network that supplies the interaction cross section used to weight the events generated by the GAN. Beyond a prediction at a fixed neutrino energy, it must return the cross section averaged over an energy interval, so that flux-averaged cross sections can be evaluated over the measurement bins. We describe in turn the tabulated training data, the input and target parametrization, the architecture, and the optimization.
Training data and interval averaging
We first tabulate cross-section values with NuWro, sampling neutrino energies from 0 to 20000 MeV in steps of 5 MeV,
with the corresponding cross sections
in units of . Below the reaction threshold, the cross section is set to zero. To obtain the average cross section over an energy interval (a bin width, which in our analysis reads 50 MeV), two table entries with indices , , corresponding to , and , , are selected at random, and the cross section is averaged over all tabulated values within the interval,
which is the flux-averaged cross section under the assumption of a uniform flux across the interval; this local-uniformity assumption is what motivates the finer flux subdivision used in following sections. The single-energy prediction is recovered as the special case .
Input and target parametrization
Since the cross section is unbounded, growing approximately linearly at large neutrino energies, the network is trained not on directly but on the bounded target
where dividing by bounds the quantity from above, and after rescaling is in GeV units. The physical cross section is recovered by inverting Eq. S-II.4 after prediction. The energy inputs are rescaled to the unit interval,
and the input for each sample is the tuple
To improve convergence, the interval width is constrained during training by MeV, so the model is trained only on intervals narrower than 1 GeV.
Architecture
The network is built from seven blocks, each consisting of a dense layer followed by layer normalization and a hyperbolic-tangent activation. The blocks contain 512, 256, 128, 64, 32, 16, and 8 neurons, respectively. A final dense layer produces a single output, activated by a softplus function
whose lower bound guarantees a non-negative predicted cross section; the factor of 10 controls the sharpness of the softplus knee.
Optimization
Three choices act together to keep the training target strictly positive. First, the softplus output cannot be negative. Second, to avoid zeros in the tabulated cross sections, we apply
with . Third, the loss is normalized by the target itself,
where is the network prediction. This target-normalized form prevents the large-cross-section region from dominating the loss, so that the low-energy, near-threshold region, where is small, remains well fit; it also requires , which the two preceding steps ensure. We train with the Nadam optimizer [33] at a learning rate of 0.001, using mini-batches of size . The single-energy result () is shown in Figure S2.
Figure S2. Cross section predicted by NuWro and by the base neural network. Predictions were made on the energies sampled for training (NuWro) and points in the middle of the energies of the tables (DNN).
The trained -DNN is subsequently used as a flux reweighter. For each flux-energy bin it provides the cross-section average defined in Eq. (S-II.3), which determines the event-energy sampling probability used in the main text. Figure S2 validates the resulting single-energy prediction against the NuWro table.
Generative adversarial network: architecture and training
A generative adversarial network (GAN) consists of two networks trained in competition: a generator and a discriminator . The discriminator learns to distinguish real samples, drawn from the training set, from fake samples produced by , while learns to produce samples that judges to be real. At convergence, produces samples indistinguishable from the real ones. GANs for scattering and their domain adaptation via transfer learning were discussed in Refs. [17, 18]. We use a conditional GAN (cGAN), since we reproduce the event kinematics as a function of the neutrino energy .
Implementation and building blocks
We implement the network in Keras 2.15 [34]. The basic feature-extraction block combines a dense layer, layer normalization [35], an activation, and a dropout layer. The activations are ReLU [36] for and PReLU [37] for . Zero-centered Gaussian noise with learnable parameters is injected into the inputs and after each layer (except for the first dense layer that receives the noisy inputs), at a rate of 0.0001 for and 0.001 for .
The generator receives a 50-dimensional latent vector drawn from together with the rescaled neutrino energy. Each of the two inputs is processed by its own dense layer with 100 neurons, then concatenated, and dropout is applied, and passed to the main body of . The generator outputs a pair of values representing the muon kinematic variables, activated by a hyperbolic tangent that constrains them to . The discriminator receives the tuple together with , from either the training set or a generated sample, which are concatenated and passed to a dense layer with 282 neurons, then dropout, and finally to the main body of . The main body of comprises three consecutive blocks of 100 neurons each; the body of comprises six blocks of 141 neurons each, so that each block carries roughly twice the number of parameters of a block.
Variable parametrization
The networks do not process the physical variables , , directly. Instead, we use the transformations
with GeV (the lowest neutrino energy in the training NuWro dataset is GeV) and is the base of the natural logarithm. Training on the muon kinematics proved difficult, especially at low energies GeV, which motivates this choice: the parametrization (empirically tuned to flatten the distribution near the phase-space boundaries) constrains the outputs to physically valid events; the muon can be neither more energetic than the incoming neutrino nor less energetic than its rest mass, and it pushes the kinematic distributions away from the edges of the allowed phase space. As a result, need not drive its weights to large values to reproduce the distribution, while can still resolve its tails.
Dense skip connections
Because adversarial training is inherently unstable, we make heavy use of skip connections to ease the flow of information through both networks. The output of each block is made available as an input to every subsequent block, yielding densely connected networks; the inputs to each block are averaged over their number, as shown in Figure S3. Such architectures have been explored extensively, e.g. residual networks [38] and densely connected networks [39].
Figure S3. The output of each block of and is used as an input to all subsequent blocks; all inputs to a block are averaged before it is processed.
Training
The GAN was trained from scratch on a CC –argon sample generated with NuWro, to reproduce the muon energy and scattering angle distributions. The training set contains 20 million events with neutrino energies in the range –20000 MeV. The spectrum is deliberately irregular, with more events generated at low energies than at high ones, and the same spectrum is used to feed ; this prevents from learning the energy spectrum instead of the muon kinematics.
Each discriminator batch consists of 2000 training events plus 2000 from , while each generator batch consists of 4000 events; one generator step is taken for every five discriminator steps. The model was trained for several hundred epochs. The learning rate was initialized at for D and for G and reduced by a factor of ten in stages over the course of training. The optimizer was AdamW with amsgrad [40] enabled at the final stages of training.
Fine-tuning
Reservoir construction and the regularization sub-step
The fine-tuning must reconcile two mismatched representations: the GAN operates on individual events, while the MicroBooNE result [23] is a three-dimensional differential cross section in bins of . To supply the GAN with event-by-event information informed by the measurement, we construct a reservoir: a sample of NuWro events whose target occupancies in are determined from the measured cross section and the MicroBooNE neutrino flux.
From the measured cross section to reservoir occupancies
The MicroBooNE measurement is reported as a smeared cross section , related to a truth-level prediction by
where is the smearing matrix provided with the data, acting on the measurement bins.
Let denote the measured double-differential cross section in the neutrino-energy region , muon-momentum bin , and angular bin . To construct non-negative sampling probabilities for the reservoir, we define the bin weight
The absolute value is used only in constructing the reservoir because the regularized unfolding yields negative central values in some low-statistics bins, which cannot correspond to negative event counts. The original signed measurement is retained in the subsequent covariance-based fit.
For each neutrino-energy region, we define the corresponding integrated non-negative weight
Let
denote the fraction of the MicroBooNE flux within contained in region , such that
The relative population of neutrino-energy regions in the reservoir is then
Within a fixed neutrino-energy region, the measured cross section determines the relative population of the muon-kinematic bins,
Consequently, the probability assigned to a three-dimensional reservoir bin is
For a reservoir containing events, the target occupancy of each bin is therefore .
Drawing the reservoir
We start from a large NuWro sample generated with the MicroBooNE neutrino-energy spectrum. For each three-dimensional bin , NuWro events falling in that bin are selected according to the target occupancy defined by Eq. (S-IV.7). The selected events are then combined and shuffled to form the one-million-event reservoir. The reservoir therefore reproduces the flux-weighted event occupancies inferred from the measured cross section while retaining truth-level NuWro kinematics.
The reservoir is not a direct experimental target. In particular, the measurement used to construct it is smeared, whereas the selected NuWro events retain their truth-level kinematics. The reservoir therefore acts as a data-informed regularization sample; the quantitative fit to the measurement is performed separately in the second sub-step.
Figure S4. One-million-event reservoir used to fine-tune the GAN. The reservoir is constructed by resampling truth-level NuWro events according to flux-weighted target occupancies inferred from the MicroBooNE measurement. The visible bin structures reflect the experimental binning.
During base training, the real and generated samples are constructed with the same spectrum, preventing the discriminator from learning to distinguish them from the neutrino-energy distribution alone. During fine-tuning, the generator instead samples from the flux reweighted by the current -DNN prediction. Only the final layers of both the generator and discriminator are updated, while the preceding layers remain fixed. The -DNN is also partially frozen, with only its final nine Keras layers left trainable during fine-tuning. This preserves most of the representation learned during base training while allowing the generated kinematics to adapt to the data-informed reservoir.
During fine-tuning, the discriminator is updated on every reservoir batch. Every fifth discriminator update is followed by an adversarial update of the generator and then by the update of the generator and the -DNN.
Role of the reservoir
An important subtlety concerns what the reservoir represents. Its target occupancies are derived from the smeared measurement and the MicroBooNE flux, whereas the reservoir itself consists of truth-level NuWro events. It therefore cannot be interpreted as a truth-level estimate of the measurement: the smearing matrix is not inverted at any point, and such an inversion would in any case be ill-posed given the regularized unfolding.
In the adversarial sub-step, the reservoir provides the real-event sample for the update of the discriminator and, through the discriminator, of the generator. Unlike the objective, the adversarial loss does not compare binned distributions; rather, it constrains the generated event distribution implicitly through the discriminator. In this way, the reservoir regularizes the generator toward a data-informed but physically valid, NuWro-like manifold of events and helps prevent mode collapse.
The quantitative comparison with the data is carried out in the second sub-step: the generated events are histogrammed, normalized using the -DNN prediction, forward-smeared by , and compared with the original signed measurement through the objective. Because this comparison is performed directly in the measurement space, it constrains both the shape and the normalization.
The two sub-steps thus play complementary roles. Without the first sub-step, the term alone reproduces the smeared, flux-averaged distributions but distorts the QE and -resonance structure (Figure 3 of the main text), which the smearing and the coarse binning render invisible to . The reservoir supplies this event-level differential structure, while the term performs the actual fit to the MicroBooNE data.
Second sub-step: differentiable matching to the differential cross section
In the second sub-step the generator and the -DNN are updated together so that the generated, cross-section-weighted, forward-smeared distribution matches the MicroBooNE differential measurement. The central difficulty is that this matching requires histogramming the generated events, which is not differentiable; we resolve it with a smooth histogram approximation described below.
Flux subdivision and the flux-averaged cross section
The measurement bins are generally wider than, and not aligned with, the flux bins . We therefore subdivide each bin into finer sub-bins of width , preserving the flux normalization within each parent bin, so that every measurement region contains an integer number of bins (Figure S5). For a region the flux-averaged cross section can then be obtained in two equivalent ways,
Figure S5. Left: A representation of a flux bin of width divided into sub-bins of flux of width . Right: Measurement bins of width that may contain several bins or fractions, but always entire bins.
the first using only the -DNN output , the second relying on the sampled events ( of falling in ). The two expressions are analytically equivalent. In the optimization, we use the “prediction” expression based directly on the -DNN output, while the “MC” expression makes explicit the equivalent relation between the flux-averaged cross section and the fraction of generated events falling in the region .
Differentiable histogram
The generator output is first mapped to the physical variables . To histogram these events differentiably, we replace the hard assignment of an event to a bin by a smooth one built from sigmoids [41], . For a bin with edges and width , the contribution of the -th event, with value , is the difference of two sigmoids,
where ( = 100) controls the sharpness of the bin edges. The first sigmoid turns on once passes the left edge, the second once it passes the right edge; their difference is for an event inside the bin and outside, recovering hard binning as while remaining smooth and differentiable for finite . The soft bin count is
and for the two-dimensional distribution the joint bin is the product
Because is a smooth function of the event positions, and these depend on the generator weights, the gradient of the loss propagates through the histogram back to ; a hard histogram would give a vanishing gradient and no training signal. This is the standard soft-histogram construction [41]; its accuracy to exact binning at our chosen is shown in Figure S6. The soft count is an estimator of the exact count; events near a bin edge are shared between neighbors, negligible at large but validated in Figure S6.
Figure S6. Left: double differential cross section as a function of from the base DNN models, using exact event counts to create histograms, and the result using the differentiable histogram approximation, with a batch of 4000 events. Right: same data as left, but smeared with the matrix .
From histogram to smeared prediction
The soft histogram is reweighted by the flux-averaged cross section to form an estimate of the double-differential cross section in region ,
where is the number of generated events with in the energy interval . Eq. (S-IV.12) is then forward-smeared by the additional smearing matrix, , bringing the prediction into the same space as the published measurement. The smearing is applied to the prediction and never inverted on the data, so the comparison is well-posed.
Gradient flow and the role of the -DNN
The -DNN enters this sub-step in two distinct roles. As the reweighter that defines the sampling distribution of , it is treated as a fixed input and does not contribute to the gradient. As the predictor of the flux-averaged cross section that rescales the histogram [Eq. S-IV.12], it is differentiated, and this is the path through which its weights are updated. Thus the sampling of carries no gradient, while the cross-section normalization does.
Loss
The generator and -DNN weights are updated by minimizing the full-covariance ,
where is the measured smeared cross section, the smeared prediction of Eq. S-IV.12, the measurement covariance, and index the flattened bins. Additionally, as a complementary diagnostic, we also report the diagonal metric
where is the variance of .
Training results
In Figure S7, we repeat Figure 2 with additional plots of (MC−Data)/Uncertainty to show the quality of our model’s predictions.
Figure S7. Base and tuned models compared with MicroBooNE two-dimensional differential cross-section data in Bins [23] on the left, and the (MC−Data)/Unc ratio on the right. The black plot in the ratio figure shows Data−|Data| to highlight the negative cross-section values found in the experimental measurement, and the gray bands show the measurement uncertainties. Grey vertical dashed lines show the different regions.
Comparison with out-of-training MicroBooNE measurements
In Table S1, we present the values for comparison with MicroBooNE measurements [24] for , and cross sections. These numbers refer to Figure 4.
| Base | 58.1/10 | 54.2/10 | 29.1/11 | 69.1/11 | 17.1/8 | 41.3/8 |
| Tuned | 44.2/10 | 0.5/10 | 12.7/11 | 6.4/11 | 4.4/8 | 6.0/8 |
Table S1. and over the number of bins of the base and tuned DNN models for , , and .
In Figure S8, we compare our models’ predictions for the and cross-section distributions. Table S2 gives the and values for the base and tuned models.
Figure S8. Base model, tuned model, and MicroBooNE data [25, 26] for (left) and (right).
| Base | 257.0/69 | 554.7/69 | 291.5/17 | 2991.7/17 |
| Tuned | 109.2/69 | 87.9/69 | 6.8/17 | 4.8/17 |
Table S2. Same as Table S1 but for and .
ArgoNeuT results for neutrinos in neutrino mode.
In Figure S9 we present both distributions of cross sections reported by ArgoNeuT [27, 28], namely and . Table S3 compares the base and tuned models to the ArgoNeuT and measurements.
Figure S9. Base model, tuned model, and ArgoNeuT data [27, 28]. Left: . Right: . Results for neutrinos in the NuMI neutrino beam. The shaded areas on the right-side plot show the range of the training MicroBooNE data.
| Base | 13.2/14 | 6.1/11 |
| Tuned | 11.4/14 | 6.0/11 |
Table S3. Comparison of MicroBooNE-tuned DNN to ArgoNeuT measurement of muon neutrino on NuMI neutrino beam.
ArgoNeuT results for neutrinos in antineutrino mode
In this section, we compare predictions from our system with ArgoNeuT measurements for neutrinos in an antineutrino beam. Table S4 presents the values computed for fine-tuned and base models. In Figure S10 we show the distributions of and cross sections and predictions of fine-tuned and base models.
| Base | 9.1/18 | 9.6/20 |
| Tuned | 9.6/18 | 9.1/20 |
Table S4. Comparison of MicroBooNE-tuned DNN to ArgoNeuT measurement of muon neutrino on NuMI anti-neutrino beam.
Figure S10. Base model, tuned model, and ArgoNeuT data [27, 28]. Left: . Right: . Results for neutrinos on the NuMI antineutrino beam. The shaded areas on the right-side plots show the range of the training MicroBooNE data.
References
References
- [11]Y. Hayato and I. Pickering, Eur. Phys. J. ST 230, 4469 (2021), arXiv:2106.15809 [hep-ph].DOI
- [12]J. Isaacson, W. I. Jay, A. Lovato, P. A. N. Machado, and N. Rocco, Phys. Rev. D 107, 033007 (2023), arXiv:2205.06378 [hep-ph].DOI
- [13]J. M. Campbell et al., SciPost Phys. 16, 130 (2024), arXiv:2203.11110 [hep-ph].DOI
- [14]S. Pina-Otey, V. Gaitan, F. Sánchez, and T. Lux, Phys. Rev. D 102, 013003 (2020), arXiv:2005.12719 [hep-ex].
- [15]M. El Baz, F. Sánchez, N. Jachowicz, K. Niewczas, A. K. Jha, and A. Nikolakopoulos, Phys. Rev. D 111, 113001 (2025), arXiv:2502.14452 [hep-ex].arxiv.org/abs/2502.14452
- [16]Y. Alanazi et al., Phys. Rev. D 106, 096002 (2022), arXiv:2008.03151 [hep-ph].arxiv.org/abs/2008.03151
- [17]J. L. Bonilla, K. M. Graczyk, A. M. Ankowski, R. D. Banerjee, B. E. Kowal, H. Prasad, and J. T. Sobczyk, Phys. Rev. D 112, 013007 (2025).
- [18]J. L. Bonilla, K. M. Graczyk, A. M. Ankowski, R. D. Banerjee, B. E. Kowal, H. Prasad, and J. T. Sobczyk, Phys. Rev. D 113, 053001 (2026), arXiv:2508.12987 [hep-ph].DOI
- [19]K. M. Graczyk, B. E. Kowal, A. M. Ankowski, R. D. Banerjee, J. L. Bonilla, H. Prasad, and J. T. Sobczyk, Phys. Rev. Lett. 135, 052502 (2025).
- [20]K. Abe et al. (T2K), Nature 580, 339 (2020), [Erratum: Nature 583, E16 (2020)], arXiv:1910.03887 [hep-ex].DOI
- [21]A. Andreassen, P. T. Komiske, E. M. Metodiev, B. Nachman, and J. Thaler, Phys. Rev. Lett. 124, 182001 (2020), arXiv:1911.09107 [hep-ph].DOI
- [22]P. Abratenko et al. (MicroBooNE), Phys. Rev. D 105, 072001 (2022), arXiv:2110.14028 [hep-ex].DOI
- [23]P. Abratenko et al. (MicroBooNE), Phys. Lett. B 870, 139939 (2025), arXiv:2307.06413 [hep-ex].DOI
- [24]P. Abratenko et al. (MicroBooNE), Phys. Rev. Lett. 128, 151801 (2022), arXiv:2110.14023 [hep-ex].DOI
- [25]P. Abratenko et al. (MicroBooNE), Phys. Rev. Lett. 133, 041801 (2024), arXiv:2402.19281 [hep-ex].DOI
- [26]P. Abratenko et al. (MicroBooNE), Phys. Rev. D 110, 013006 (2024), arXiv:2402.19216 [hep-ex].DOI
- [27]C. Anderson et al. (ArgoNeuT), Phys. Rev. Lett. 108, 161802 (2012), arXiv:1111.0103 [hep-ex].DOI
- [28]R. Acciarri et al. (ArgoNeuT), Phys. Rev. D 89, 112003 (2014), arXiv:1404.4809 [hep-ex].DOI
- [29]See Supplemental Material for additional details on the cross-section DNN, GAN architecture and training, fine-tuning procedure, differentiable histogramming, and comparisons with out-of-training MicroBooNE and ArgoNeuT measurements.
- [30]S. Abe et al., Phys. Rev. D 113, 012011 (2026), arXiv:2509.17945 [hep-ex].DOI
- [31]P. Adamson et al., Nucl. Instrum. Meth. A 806, 279 (2016), arXiv:1507.06690 [physics.acc-ph].DOI
- [32]J. L. Bonilla and K. M. Graczyk, Numugen, https://github.com/LuisBonillaR/NuMuGen (2026).
- [33]T. Dozat, in Proceedings of the 4th International Conference on Learning Representations, Workshop Track (San Juan, Puerto Rico, 2016) pp. 1–4.
- [34]F. Chollet et al., Keras, https://keras.io (2015).
- [35]J. L. Ba, J. R. Kiros, and G. E. Hinton, Layer normalization (2016), arXiv:1607.06450 [stat.ML].arxiv.org/abs/1607.06450
- [36]X. Glorot, A. Bordes, and Y. Bengio, in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, Vol. 15, edited by G. Gordon, D. Dunson, and M. Dudík (PMLR, Fort Lauderdale, FL, USA, 2011) pp. 315–323.
- [37]K. He, X. Zhang, S. Ren, and J. Sun, arXiv e-prints , arXiv:1502.01852 (2015), arXiv:1502.01852 [cs.CV].
- [38]K. He, X. Zhang, S. Ren, and J. Sun, CoRR abs/1512.03385 (2015), 1512.03385.
- [39]G. Huang, Z. Liu, and K. Q. Weinberger, CoRR abs/1608.06993 (2016), 1608.06993.
- [40]S. J. Reddi, S. Kale, and S. Kumar, CoRR abs/1904.09237 (2019), 1904.09237.
- [41]M. Avi-Aharon, A. Arbelle, and T. R. Raviv, Deephist: Differentiable joint and color histogram layers for image-to-image translation (2020), arXiv:2005.03995 [eess.IV].











