AI takes on one of physics’ biggest mysteries
Observing a neutrino is a lot like listening to a weak radio signal. It’s there, but it can be hard to pinpoint and decipher what the signal is saying. A neutrino is a tiny, nearly invisible particle that rarely interacts with matter. When it does, it leaves behind a signal called a neutrino event. During experiments in a particle accelerator, millions of these events can occur, and researchers need to decipher each one.
Classifying millions of neutrino events by hand isn’t practical and would take an enormous amount of time. Now, a research team from UC Irvine is using advances in machine learning to classify these events and better understand one of the most puzzling phenomena in the physical world.
The elusive particle
Neutrinos are the most common particle with mass in the universe. Nevertheless, they are hard to observe, and they can inexplicably transform as they travel through space. This transformation or change in “flavor” is a process called “neutrino oscillation.”
The problem? This behavior isn’t fully explained by the Standard Model – the reigning theory used to explain the fundamental building blocks of the universe.
To study this transformation, scientists rely on identifying the type of neutrino interaction in detector data. When a neutrino interacts with matter, it can produce different patterns depending on the nature of the interaction. Each neutrino event offers scientists insight into the mysterious particle. The three event classifications are muon neutrino, which produces a long, straight track; electron neutrino, which produces a fuzzy, shower-like pattern; and neutral current, which is relatively messier and leaves no clear track.
By knowing the type of neutrino you started with and the types of events that occurred in the particle detector, researchers can then infer neutrino oscillation.
To better meet that challenge, researchers are turning to artificial intelligence.
Classifying neutrino events
Traditional approaches rely on machine learning models called convolutional neural networks, or CNNs. These models analyze images of neutrino events by learning small patterns and combining them into larger ones, much like moving a magnifying glass across a map. While effective, these models don’t provide insight into their reasoning.
The research team took a different approach – one that combines visual analysis with language-based reasoning.
In a recent paper published in Nature Communications Physics, the team used simulated data from a liquid argon time projection chamber to test multimodal language models’ ability to accurately classify neutrino events.
Instead of using standard image-based models, the group adapted a vision-language model, or VLM – a type of artificial intelligence that can analyze images and describe them in words. Similar to tools like ChatGPT that can interpret images and explain what they see, these models combine visual recognition with written reasoning.
Researchers used a vision-transformer model, or ViT, as an intermediate baseline between VLM and CNN because it has a more advanced architecture than CNNs but is not as advanced as a VLM. A standalone ViT model does not generate language-based explanations for its predictions.
By fine-tuning a vision-language model on neutrino detector data, the researchers trained it not only to classify events, but also to explain its decisions.
Beyond accuracy
The results showed clear improvement over traditional methods.
But accuracy isn’t the only advantage. VLM also offers interpretability and improved flexibility in how the model is trained.
Instead of simply labeling an event, the model can generate explanations such as identifying a “long, narrow track” consistent with a muon neutrino or a “diffuse shower” that’s characteristic of an electron neutrino.
For scientists, this added layer of explanation is critical.
“It’s not just about getting the right answer,” said Jianming Bian, professor of physics & astronomy. “It’s about understanding why and enabling scientists to communicate with AI through a shared language of reasoning.”
The VLM also has a practical edge in retraining. Certain modern AI models like contain billions of parameters, making full retraining slow, expensive and prone to erasing what the model already knows. Using a technique called parameter-efficient fine-tuning, researchers can freeze most of the base model and update only a small set of parameters. Think of it like an expert scientist attending a workshop instead of redoing school.
“The next step is to improve the model’s ability to explain its reasoning by incorporating feedback from students and scientists. As these systems become better at communicating physics concepts and scientific reasoning, they could serve not only as research tools for neutrino experiments, but also as educational tools that help train the next generation of physicists.” Bian said.
This work demonstrates that multimodal AI can outperform traditional methods in scientific data analysis while adding the crucial capability of human-readable explanations, making it a promising direction for future physics research.
“AI is rapidly expanding to all areas of physics, not just neutrino physics,” said Pierre Baldi, Distinguished Professor of computer science and founding director of the UC Irvine AI in Science Institute. “For instance, we are also using AI to address problems in astronomy. One of the most intriguing questions is whether AI will be able to make significant contributions to theoretical physics.”