Transfer Learning for Neutrino Scattering: Domain Adaptation with GANs

This paper demonstrates that transfer learning with Generative Adversarial Networks effectively extrapolates physics information from synthetic neutrino-carbon scattering data to related processes like neutrino-argon and antineutrino-carbon interactions, significantly outperforming models trained from scratch and maintaining high accuracy even with limited statistics.

The Gist

The Big Idea: Teaching a Robot to "Guess" Physics

Imagine you are trying to teach a robot how to predict what happens when a tiny particle (a neutrino) smashes into an atom. This is crucial for experiments like DUNE and Hyper-Kamiokande, which are trying to solve mysteries about the universe.

The problem? We don't have enough real data. Neutrinos are ghosts; they rarely hit anything, so getting millions of real-life collision records is incredibly hard and expensive.

Usually, scientists use complex math simulations (like a video game engine called NuWro) to create fake data to train their models. But these simulations aren't perfect. They are like a map drawn by someone who has never actually been to the city—they get the general layout right, but the street names might be wrong.

This paper introduces a clever trick called Transfer Learning (TL) using GANs (Generative Adversarial Networks). Think of a GAN as a team of two robots:

1. The Forger: Tries to create fake particle collision data that looks real.
2. The Detective: Tries to spot the difference between the fake data and the real (or simulated) data.

They play a game of cat-and-mouse. The Forger gets better at faking it, and the Detective gets better at spotting fakes. Eventually, the Forger becomes so good that it can generate perfect physics data.

The Problem: Starting from Scratch is Hard

In the past, if scientists wanted to study a new type of collision (say, a neutrino hitting an Argon atom instead of a Carbon atom), they had to train a brand-new "Forger" from zero.

Imagine you are learning to play the piano. If you want to learn a new song, you have to start from the very first note, even if you already know how to play scales. If you only have a few hours to practice (limited data), you'll probably sound terrible.

In physics, this means if you have very few experimental data points for Argon, a model trained from scratch will fail to understand the complex patterns of the collision.

The Solution: The "Musical Prodigy" Analogy

The authors asked: What if we took a Forger that was already an expert at playing "Carbon" collisions and just taught it the new "Argon" song?

This is Transfer Learning.

1. The Pre-trained Model (The Prodigy): They started with a GAN that had already mastered simulating neutrinos hitting Carbon atoms. This model had already learned the "universal rules" of how neutrinos behave—like how they bounce off nuclei, where the energy peaks are, and the general shape of the collision.
2. The Fine-Tuning (The Lesson): Instead of retraining the whole robot, they "froze" the part of the brain that knows the universal rules and only retrained the part that handles the specific details (like the size of the Argon atom).
3. The Result: They gave this "Carbon expert" a tiny amount of Argon data (as little as 10,000 events, which is very small in physics terms). Because the robot already knew the basics, it only needed to learn the specific differences.

The Three Tests

The team tested this "Carbon expert" on three new scenarios to see if it could adapt:

• Scenario A: The Heavyweight (Argon): They asked the model to simulate neutrinos hitting Argon. Argon is much heavier and more complex than Carbon.
  • Result: The Transfer Learning model nailed it. A model trained from scratch struggled and sounded like a beginner. The "expert" just needed a quick refresher.
• Scenario B: The Opposite Charge (Antineutrinos): They asked the model to simulate antineutrinos hitting Carbon. Antineutrinos are like the "evil twins" of neutrinos; they interact slightly differently.
  • Result: Again, the Transfer Learning model was far superior. It understood the core physics and just adjusted for the "evil twin" behavior.
• Scenario C: The New Rulebook (Different Simulation): They used a different version of the simulation software (NuWro) with updated physics rules.
  • Result: Even when the underlying "textbook" changed, the Transfer Learning model adapted quickly, while the "from scratch" model got confused.

Why This Matters

Think of it like this:

• Training from scratch is like trying to build a house by mining your own sand and making your own bricks. It takes forever and requires a massive pile of resources.
• Transfer Learning is like buying a pre-fabricated house frame that is already 90% built. You just need to add the paint and the specific furniture for your family.

The Takeaway:
This paper proves that we don't need millions of data points to build accurate physics models. By using a "pre-trained" AI that understands the universal laws of neutrino physics, we can quickly and accurately adapt it to new targets (like Argon) or new conditions, even when we have very little data.

This is a game-changer for future experiments. It means scientists can build better "event generators" (the tools that predict what will happen in their detectors) faster and more accurately, even when the universe is being stingy with data.

Technical Summary

1. Problem Statement

Simulating (anti)neutrino-nucleus interactions is critical for next-generation neutrino oscillation experiments (e.g., Hyper-Kamiokande, DUNE). Traditional Monte Carlo (MC) event generators rely on theoretical models with adjustable parameters tuned to sparse experimental data. However, these models often suffer from fundamental limitations in nuclear physics approximations (e.g., spectral functions vs. local Fermi gas).

Recent advances in Generative Adversarial Networks (GANs) offer a data-driven alternative to simulate these interactions. However, training GANs from scratch requires massive datasets, which are often unavailable for specific target nuclei (e.g., Argon) or specific interaction channels (e.g., antineutrinos). The core problem addressed is: Can a GAN trained on one domain (e.g., neutrino-carbon scattering) be effectively adapted to generate accurate simulations for related but distinct domains (e.g., neutrino-argon, antineutrino-carbon) with limited data?

2. Methodology

A. Transfer Learning (TL) and Domain Adaptation

The authors employ Transfer Learning, specifically Domain Adaptation, to leverage a pre-trained GAN model.

• Source Domain: A GAN pre-trained on synthetic inclusive charged-current (CC) $\nu_\mu$-carbon ($^{12}\text{C}$) scattering data generated by the NuWro MC generator (version 21.09).
• Target Domains: The pre-trained model is fine-tuned to simulate three distinct scenarios:
  1. $\nu_\mu$-Argon ($^{40}\text{Ar}$): A different target nucleus with a different spectral function (SF) and no optical potential in the SF model used.
  2. $\bar{\nu}_\mu$-Carbon: Antineutrino scattering on the same target, involving different vector/axial current interference.
  3. $\nu_\mu$-Carbon (Alternative Model): Scattering on carbon but using a different NuWro configuration (version 25.03) with modified axial masses and a Local Fermi Gas (LFG) model instead of the Spectral Function.

B. GAN Architecture

The study utilizes a modified architecture based on the authors' previous work [23]:

• Generator: Takes a random latent vector (Gaussian) and normalized neutrino energy ($E'_\nu$) as input. It outputs proxy kinematic variables for the muon: normalized energy ($E'_\mu$) and scattering angle ($\theta'$).
  • Modification: The generator now injects Gaussian noise after every learnable layer (previously only in the discriminator). The architecture uses "Block 3" and "Block 4" (which include energy conditioning) instead of the original Block 1/2.
• Discriminator: Takes the proxy variables ($v = \{E'_\mu, \theta'\}$) and neutrino energy as input to distinguish real (NuWro) from fake (GAN) events.
• Freezing Strategy: During fine-tuning, the first block of both the Generator and Discriminator is frozen. This block is hypothesized to encode universal kinematic constraints. The remaining blocks are unfrozen and fine-tuned on the target domain data.

C. Training and Metrics

• Data: Training datasets of 10,000 and 100,000 events were used for fine-tuning to simulate both sparse and abundant data scenarios.
• Loss Function: The authors propose a hybrid loss function for the generator, combining the standard cross-entropy loss and the non-saturating heuristic loss. This was found to prevent discriminator saturation and improve performance across low and high neutrino energies.
• Evaluation Metrics:
  • Earth Mover's Distance (EMD): Measures the distance between probability distributions.
  • Mean Averaged Pull (MAP-3D): A 3D metric evaluating the agreement between GAN and NuWro distributions in the space of ($E'_\nu, E'_\mu, \theta'$).

3. Key Contributions

1. Validation of Universal Kinematic Features: The study demonstrates that deep generative models capture universal features of lepton-nucleus dynamics (kinematic constraints, QE and $\Delta(1232)$ resonance patterns) that are transferable across different nuclei and interaction types.
2. Efficiency in Data-Scarce Regimes: The paper proves that TL allows for high-accuracy model generation with as few as 10,000 events, a regime where models trained from scratch fail to reproduce physical structures (like resonance peaks).
3. Robustness to Model Mismatch: The TL approach successfully adapts a model trained on one theoretical framework (NuWro v21.09 with SF) to a significantly different framework (NuWro v25.03 with LFG and different axial masses), correcting target-dependent spectral details through fine-tuning.
4. Architectural Insight: By freezing the initial layers, the authors show that the first block of the network acts as a universal encoder for lepton kinematics, while subsequent layers handle target-specific spectral details.

4. Results

• Performance Comparison:
  • Small Data (10k events): TL models significantly outperformed "from-scratch" models. For $\nu_\mu$-Argon, the TL model achieved an EMD of 0.36 compared to 0.96 for the scratch model. The scratch model failed to reproduce the Quasielastic (QE) and $\Delta$-resonance peaks, while the TL model captured them accurately.
  • Large Data (100k events): While scratch models improved with more data, TL models consistently maintained lower error rates (better MAP and EMD scores) and converged faster.
• Specific Scenarios:
  • $\nu_\mu$-Argon: TL successfully adapted the carbon model to the heavier argon nucleus, correcting the broader QE peak caused by the argon spectral function.
  • $\bar{\nu}_\mu$-Carbon: TL effectively handled the sign change in vector/axial contributions, reproducing the suppressed total cross-section and altered peak shapes.
  • Alternative Model: TL adapted to the LFG model and modified axial masses, correcting the sharper QE peak and resonance structures inherent to the new theoretical parameters.
• Visual Evidence: Histograms of reconstructed hadronic invariant mass ($W$) showed that TL models accurately reproduced the QE and $\Delta(1232)$ resonance structures, whereas scratch models often smoothed over or missed these features entirely in low-statistics regimes.

5. Significance

• Accelerating Event Generator Development: This work provides a computationally efficient pathway to build next-generation neutrino event generators. Instead of training massive models for every new target or interaction type, physicists can start with a robust, pre-trained "universal" model and fine-tune it with limited experimental or synthetic data.
• Bridging Theory and Experiment: The method is particularly valuable for experiments like DUNE (using Argon) where experimental data is currently sparse. It allows for the construction of realistic simulators based on theoretical priors that can be rapidly updated as new data arrives.
• Physical Interpretability: The success of the "frozen first layer" approach suggests that a significant portion of neutrino-nucleus interaction physics is universal. This offers new insights into which aspects of lepton-hadron dynamics are fundamental and which are specific to nuclear structure.
• Future Directions: The authors plan to extend this framework to full final-state hadron generation (currently limited to inclusive muon kinematics) and apply it to real experimental flux-averaged data.

In conclusion, the paper establishes Transfer Learning with GANs as a powerful, physically motivated tool for simulating neutrino interactions, offering superior accuracy and efficiency over traditional training methods, especially in the data-limited regimes characteristic of modern neutrino physics.