⚛️ high-energy experiments

Machine-learning-based neutrino-event generation constrained by experimental measurements

This paper presents a machine-learning-based generative surrogate for charged-current neutrino-nucleus scattering that can be directly calibrated to experimental differential cross sections without event-level data, thereby improving predictions for unmeasured observables and maintaining applicability across different neutrino fluxes.

Original authors: Jose Luis Bonilla, Krzysztof M. Graczyk, Rwik Dharmapal Banerjee, Beata E. Kowal, Hemant Prasad, Jan T. Sobczyk

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jose Luis Bonilla, Krzysztof M. Graczyk, Rwik Dharmapal Banerjee, Beata E. Kowal, Hemant Prasad, Jan T. Sobczyk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand the universe, physicists often look at the smallest things that make it up. Among these, neutrinos are unique ghosts: they are tiny particles that zip through almost everything without stopping, passing through the Earth, the Sun, and even our own bodies by the trillions every second. Because they interact so rarely, catching them requires massive detectors buried deep underground or under ice. When a neutrino does finally hit an atomic nucleus inside a detector, it creates a burst of other particles, including a muon, which is like a heavier cousin of the electron. By studying the speed and direction of these muons, scientists can learn about the neutrino itself, such as its energy and whether it changed its identity while traveling. This is crucial for understanding why the universe is made of matter rather than antimatter, a mystery that defines our existence. However, to read the message carried by these neutrinos, scientists must first understand exactly how they smash into atomic nuclei. This process is complex, involving quantum mechanics and the messy behavior of atoms, making it difficult to predict what happens in a collision with perfect accuracy.

For decades, researchers have relied on computer simulations to model these collisions. These programs, known as Monte Carlo generators, act as virtual laboratories where scientists can run millions of fake collisions to see what the data should look like. But these simulations are built on theoretical guesses, and if the guesses are slightly off, the entire interpretation of the real experiment can be skewed. The challenge has been that real experiments do not provide a list of every single collision that happened; instead, they provide a blurred, averaged picture of many collisions mixed together, often with different types of neutrino beams. Trying to fix the computer models to match this blurred picture has been like trying to tune a radio by listening to static.

In a new study, a team of researchers has developed a clever way to bridge this gap using artificial intelligence. They created a new kind of computer program that can learn the complex patterns of how neutrinos hit atomic nuclei, but with a special twist: it can be directly adjusted to match real experimental measurements, even when those measurements are just broad averages rather than detailed lists of individual events. The team tested their method using data from the MicroBooNE experiment, which uses a beam of neutrinos fired at a tank of liquid argon. They started with a standard computer simulation of how neutrinos interact with argon atoms. Then, they trained two separate artificial intelligence networks on this simulation. One network learned to predict the total likelihood of a collision happening at a specific energy, while the other learned to generate the specific details of what comes out of the collision, such as the speed and angle of the resulting muon.

The breakthrough came when the team combined these two networks and adjusted them to fit the actual data from MicroBooNE. Instead of trying to match every single particle track, they tuned the system to match the overall shape of the collision data, which includes how often collisions happen at different energies and angles. This process allowed the computer model to learn from the real world without needing a perfect, event-by-event record of what happened in the detector. The result was a new "surrogate" model that could generate realistic collision events that were much closer to reality than the original simulation.

The researchers then put this new model to the test in two ways. First, they checked if it could correctly predict other types of measurements from the same MicroBooNE experiment that they had not used during the training. The model succeeded, accurately describing the behavior of muons and other particles in ways that the original simulation had missed. This showed that the model had truly learned the underlying physics of the interaction, not just memorized the specific data points it was shown. Second, they tested the model on data from a completely different experiment called ArgoNeuT, which used a different type of neutrino beam and a different detector. Even though the new model had never seen this specific beam before, it still produced predictions that matched the ArgoNeuT data well. This is a significant achievement because it proves the model is not limited to one specific experiment but captures the fundamental rules of how neutrinos interact with matter.

This work demonstrates that machine learning can be used to refine our understanding of the subatomic world by directly incorporating real experimental data into the simulation process. By allowing the computer model to be corrected by the blurred, averaged data that experiments actually produce, the team has created a tool that is more reliable and versatile. This means that future experiments, which aim to measure the subtle differences between matter and antimatter with extreme precision, will have a much better foundation for their calculations. The method does not replace the need for theoretical physics but rather strengthens it, creating a hybrid system where the best of computer theory and real-world observation work together. As the next generation of neutrino experiments comes online, this approach could become a standard way to ensure that the models used to interpret the data are as accurate as possible, helping scientists unlock the deepest secrets of the universe.

Technical Summary: Machine-learning-based neutrino-event generation constrained by experimental measurements

Problem Statement
Precision measurements of neutrino oscillations and leptonic CP violation, essential for testing the Standard Model and understanding matter-antimatter asymmetry, require accurate modeling of neutrino–nucleus interactions. Nuclear effects significantly influence final states and neutrino-energy reconstruction. However, increasingly sophisticated theoretical interaction models make efficient event generation in Monte Carlo (MC) simulations challenging. Furthermore, accelerator neutrino beams possess broad, energy-dependent fluxes, meaning measured cross sections are flux-averaged and involve multiple simultaneous reaction mechanisms. This complicates model validation and the tuning of MC event generators. Existing machine-learning approaches, such as unfolding techniques (e.g., OmniFold) or parametric generator tunes, either rely on event-level detector data or adjust predefined parameters of underlying models, respectively. There is a need for a generative surrogate that can be directly calibrated to measured differential cross sections without requiring event-level experimental samples, while preserving the ability to generate event-level predictions for different beam configurations.

Methodology
The authors propose a two-stage framework combining a deep neural network (DNN) for cross-section prediction with a conditional Generative Adversarial Network (cGAN) for event generation.

Architecture:
- $\sigma$ -DNN: A feed-forward network predicts the total interaction cross section $\sigma(E_\nu)$ averaged over energy intervals. This provides the energy-dependent normalization required to weight events.
- Conditional GAN (cGAN): A generator ( $G$ ) and discriminator ( $D$ ) are trained to reproduce the kinematics of outgoing muons (energy $E_\mu$ and scattering angle $\theta$ ) conditioned on the neutrino energy $E_\nu$ . The generator maps a latent noise vector and $E_\nu$ to $(E_\mu, \theta)$ .
Training Stages:
- Stage 1 (Base Training): Both the $\sigma$ -DNN and cGAN are pre-trained on data generated by the NuWro MC event generator. The $\sigma$ -DNN learns to predict cross sections, while the cGAN learns the multidimensional distribution of muon kinematics.
- Stage 2 (Fine-tuning): The models are fine-tuned simultaneously using experimental data from MicroBooNE (inclusive charged-current $\nu_\mu$ $ν_{μ}$ -argon scattering). This stage employs a dual-optimization procedure:
  - Adversarial Regularization: A "reservoir" of one million truth-level NuWro events is constructed. The target occupancies of this reservoir are determined by the measured MicroBooNE cross sections and flux, but the events themselves retain truth-level kinematics. The discriminator is fine-tuned on this reservoir to regularize the generator, ensuring it remains on a physically valid manifold and preventing mode collapse.
  - Differential Cross-Section Fit: The generator and $\sigma$ -DNN are updated to minimize a $\chi^2_c$ loss function. This involves generating events, histogramming them using a differentiable "soft histogram" approximation (to allow gradient flow), normalizing by the $\sigma$ -DNN prediction, forward-smeared by the experimental smearing matrix $A_c$ , and comparing against the measured binned cross sections using the full covariance matrix.

Key Contributions

Direct Calibration to Binned Data: The framework enables the direct fine-tuning of a generative event model to flux-averaged, binned differential cross sections without requiring access to event-level experimental data.
Separation of Normalization and Kinematics: By coupling a cross-section predictor ( $\sigma$ -DNN) with a kinematic generator (cGAN), the model separates the energy-dependent normalization from the event structure. This allows the model to be transported to different neutrino fluxes without retuning, simply by supplying the new flux profile.
Differentiable Histogramming: The implementation of a soft-histogram approximation allows the optimization of generative models against binned experimental data, overcoming the non-differentiability of standard histogramming.
Reservoir Regularization: The use of a data-informed reservoir of truth-level events preserves the fine-grained kinematic structures (such as quasielastic and $\Delta$ -resonance peaks) that are washed out by experimental smearing and coarse binning, which the $\chi^2$ fit alone cannot constrain.

Results
The surrogate was calibrated to MicroBooNE measurements of the double-differential cross section $d^2\sigma/dp_\mu d\cos\theta$ .

In-Distribution Performance: The fine-tuned model significantly improved the agreement with MicroBooNE data compared to the base NuWro-trained model. The full-covariance $\chi^2_c$ per bin improved from 460.3 to 220.3.
Out-of-Training Observables: The tuned model improved the description of other MicroBooNE observables not used in training, including total cross sections $\sigma(E_\nu)$ , and differential distributions $d\sigma/dE_\mu$ and $d\sigma/dE_{had}$ . Notably, the model predicted a later MicroBooNE measurement of $d\sigma/d\cos\theta$ with a $\chi^2_c/n$ of 6.8/17, a substantial improvement over the base model (291.5/17), achieving a fit quality of $\chi^2_c/n < 1$ .
Cross-Experiment Consistency: The model was tested against ArgoNeuT data, which utilizes a different detector and a NuMI beam flux extending to energies well above the training range. The tuned surrogate remained consistent with ArgoNeuT data, demonstrating that the fine-tuning did not degrade the high-energy behavior inherited from the base MC model.

Significance
The paper demonstrates that experimentally measured cross sections can directly constrain generative event models while preserving their applicability across different observables and beam configurations. This approach offers a pathway to combine theory-based simulations with direct experimental constraints, creating a common interaction reference that can be folded with various beam fluxes to test the mutual consistency of cross-section measurements across different experiments. The authors note that extending the constrained kinematic range and propagating calibrated model uncertainties are natural next steps for this framework.

More like this