🔬 physics

Domain Adaptive Graph Neural Networks for Neutrino Experiments

This paper introduces the first domain-adaptation-enabled Graph Neural Networks for Liquid Argon Time Projection Chamber (LArTPC) neutrino experiments, demonstrating that adversarial training on MicroBooNE data significantly improves reconstruction robustness and performance across various simulated-to-realistic domain shifts without compromising source domain accuracy.

Original author: Hemalata Nayak, Giuseppe Cerati, Aleksandra Ćiprijanović, Tammy Walton

Published 2026-11-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Hemalata Nayak, Giuseppe Cerati, Aleksandra Ćiprijanović, Tammy Walton

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 author. For technical accuracy, refer to the original paper. Read full disclaimer

Neutrinos are ghostly particles that zip through the universe almost entirely unnoticed, passing through stars, planets, and even our own bodies without leaving a trace. Because they interact so rarely with matter, scientists must build massive, sensitive detectors to catch even a few of them. These detectors, often filled with liquid argon, act like giant 3D cameras that record the faint trails left behind when a neutrino finally bumps into an atom. To make sense of these complex trails, physicists rely on computer programs to reconstruct the event, identifying what kind of particle was involved and where it came from. In recent years, artificial intelligence has become a powerful tool for this task, helping to sort through the noise and find the signal. However, a persistent problem has plagued these efforts: the computer programs are usually trained on perfect, simulated data, but real-world detectors are messy and imperfect. When a model trained on a clean simulation is applied to the noisy reality of a physical experiment, it often stumbles, leading to errors that could hide new discoveries or create fake ones.

A team of researchers has developed a new way to teach these artificial intelligence models to handle the messiness of real life. They focused on a specific type of detector called a liquid argon time projection chamber, which is central to current and future neutrino experiments. The team used a powerful type of neural network known as a graph neural network, which is designed to understand relationships between scattered points of data, much like connecting dots to see a picture. Instead of forcing the data into a rigid grid, this network treats the detector hits as a flexible web of connections. The researchers trained this network on a large set of simulated neutrino events from the MicroBooNE experiment, which served as their "source" of truth. To test how well the model would hold up against real-world imperfections, they created four different "target" versions of the data by intentionally introducing errors. These errors mimicked common problems scientists face, such as the detector measuring energy levels slightly too high or too low, electronic interference that adds random static, or changes in how electrons drift through the liquid over time.

When they tested their standard model on these corrupted datasets, it performed poorly, struggling to distinguish between real particle signals and background noise. The model became confused by the shifts in the data, often missing valid signals or misidentifying noise as a particle. This confirmed that the gap between simulation and reality was indeed a significant hurdle. To fix this, the researchers introduced a technique called domain adaptation. They modified the training process so that the network had to learn features that were useful for identifying particles while simultaneously ignoring the specific differences between the simulated data and the altered target data. It was as if the network was forced to learn the essence of a neutrino event without getting distracted by the specific quirks of the dataset it was looking at. This was achieved by adding a special component that tried to guess whether the data came from the simulation or the altered version, while the main network fought to make that guess impossible.

The results showed that this approach worked remarkably well. When the researchers applied the adapted model to the four different corrupted datasets, its performance improved dramatically. In the most difficult scenarios, where the data was heavily distorted by noise or scaling errors, the model's ability to correctly identify signals increased significantly, recovering much of the accuracy it had lost. For example, in one test involving severe noise, the model's success rate in finding the right particles jumped from a low level back up to a high level, nearly matching its performance on the original, clean simulation. Crucially, this improvement in handling the messy data did not come at the cost of performance on the clean data; the model remained just as good at analyzing the original simulation as it was before. The researchers found that the network learned to create a shared understanding of the data that worked for both the perfect simulations and the imperfect realities.

This work demonstrates that it is possible to build more trustworthy artificial intelligence models for physics by teaching them to bridge the gap between theory and experiment. By using this domain adaptation technique, scientists can create tools that are robust enough to handle the inevitable differences between computer models and physical detectors. This is vital for the future of neutrino research, where the next generation of detectors will be even larger and more complex. If the models used to analyze data from these massive machines are not reliable, they might miss subtle signs of new physics or mistake a glitch for a discovery. The study suggests that by training models to be invariant to these specific types of errors, researchers can ensure that their conclusions are based on the actual behavior of neutrinos rather than artifacts of the simulation. While the team tested this on specific types of errors, the method offers a promising path forward for making machine learning a more reliable partner in the search for the fundamental laws of the universe.

Technical Summary: Domain Adaptive Graph Neural Networks for Neutrino Experiments

Problem Statement
Neutrino experiments, particularly those utilizing Liquid Argon Time Projection Chambers (LArTPCs) like MicroBooNE and the future DUNE, rely on highly efficient and precise reconstruction methods to detect weakly interacting particles. While Graph Neural Networks (GNNs), such as the NuGraph2 and NuGraph3 architectures, have shown promise in handling the irregular, sparse data structures of LArTPCs, they suffer from "domain shift." Models trained on simulated data often perform sub-optimally when applied to real experimental data due to discrepancies between simulation and reality (e.g., imperfect simulations, electronic noise, or calibration differences). These shifts can degrade model performance, potentially mimicking or obscuring new physics signals.

Methodology
The authors propose a Domain-Adversarial Neural Network (DANN) framework to improve the robustness of the NuGraph3 GNN against simulation-to-data discrepancies.

Architecture: The base model, NuGraph3, is a hierarchical message-passing attention network. It constructs graph representations from low-level detector hits using five node features: detector plane number, wire index, time coordinate, and pulse properties (integral and RMS). The network includes an encoder, a core message-passing block, and task-specific decoders (a filter decoder for hit rejection and a semantic decoder for particle classification).
Domain Adaptation Strategy: To mitigate domain shift, the authors integrate an adversarial branch into the training process. This branch consists of a dense layer (size 64) followed by a ReLU activation and an output layer (size 2) trained to distinguish between source (simulated) and target (perturbed) data.
Training Mechanism: A Gradient Reversal Layer (GRL) is placed before the domain classifier. During backpropagation, the GRL multiplies gradients by a negative constant, forcing the feature extractor to minimize the primary task losses (filter and semantic) on the source domain while simultaneously maximizing the error of the domain classifier. This encourages the learning of latent representations that are discriminative for physics tasks but invariant to the underlying domain.
Loss Function: The total loss is a weighted sum of the source domain filter loss ( $L^S_{filt}$ ), source domain semantic loss ( $L^S_{sem}$ ), and the domain adaptation loss ( $L^{ST}_{DA}$ ). A temperature mechanism is used to balance these multi-task losses.

Experimental Setup

Source Domain: A subset of the MicroBooNE open dataset containing simulated inclusive ( $\nu_\mu$ ) and electron neutrino ( $\nu_e$ ) interactions overlaid on real data.
Target Domains: Four synthetic datasets were generated by systematically perturbing the hit features (integral and RMS) of the source data to mimic realistic discrepancies:
1. Fixed Scaling: Independent scaling factors applied to U, V, and Y wire planes.
2. Moderate Gaussian Noise: Perturbations with a standard deviation of 20% of the global robust feature spread.
3. Severe Gaussian Noise: Perturbations with a standard deviation of 40% of the global robust feature spread.
4. Time-Dependent RMS Shift: A shift in RMS values to simulate variations in longitudinal electron diffusion constants (mimicking different simulation campaigns).
Evaluation: Models were trained with and without DA. Performance was measured using Precision and Recall on both source and target test sets.

Key Results
The study demonstrates that standard GNNs are vulnerable to domain shifts, with performance degrading significantly on perturbed target datasets.

Baseline Vulnerability: Without DA, the semantic decoder's performance dropped from ~0.89 (source) to 0.68 (fixed scaling) and 0.67 (severe noise). The filter decoder showed a sharp decline in recall (dropping to 0.56 in the fixed scaling scenario), indicating the model aggressively filtered out valid hits when exposed to out-of-distribution scales.
DA Improvement: Integrating DANN successfully recovered performance across all four target domains without substantial degradation in the source domain.
- In the severe noise scenario (DA3), semantic precision and recall rebounded from 0.67 to 0.86, and filter recall recovered from 0.80 to 0.95.
- In the fixed scaling scenario (DA1), filter recall was restored from 0.56 to 0.93.
- Precision in the target domain improved by up to 0.19, and recall by up to 0.37, depending on the variation and task.
Latent Space Analysis: UMAP visualizations of the latent embeddings show that DA improves the alignment of source and target distributions and slightly enhances the separation between signal and noise classes compared to the non-DA model.

Significance and Claims
The authors claim this work represents the first application of domain-adaptation-enabled GNNs for robust LArTPC reconstruction tasks. The primary significance lies in demonstrating that adversarial domain adaptation can effectively bridge the gap between simulations and experimental data. By training models to be invariant to specific simulation artifacts (scale variations, noise, diffusion miscalibration), the approach builds more trustworthy models for current and future neutrino experiments. The authors note that while the results are promising, future work will explore other DA methods and the application of this technique to bridge domain shifts between different LArTPC experiments with distinct detector designs.

More like this