← Latest papers
⚛️ phenomenology

Extrapolation and explainability in hadron mass predictions with machine learning

This study evaluates four machine learning models for predicting hadron masses, demonstrating that a multilayer perceptron (MLP) significantly outperforms other methods in extrapolating to unseen exotic and heavy baryon states while using SHAP analysis to identify heavy quark content as the primary predictive feature.

Original author: Esra Yuksel, Huseyin Bahtiyar, Derya Soydaner

Published 2026-11-03
📖 4 min read🧠 Deep dive

Original author: Esra Yuksel, Huseyin Bahtiyar, Derya Soydaner

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

The universe is built from a vast zoo of particles, but the most familiar members of this family are the hadrons. These are the particles that make up the protons and neutrons inside every atom, and they are held together by the strongest force in nature. For decades, physicists have been cataloging these particles, measuring their masses with extreme precision to understand how the strong force works. While we have a solid grasp of the "conventional" hadrons—those made of simple combinations of quarks—recent experiments have begun to reveal a more exotic side of the story. Scientists are now finding particles that do not fit the standard patterns, such as those containing four or five quarks, or heavy baryons packed with multiple heavy quarks. Predicting the masses of these new, strange particles is a major challenge because the theoretical tools used for ordinary particles often struggle when applied to these unusual configurations.

In a recent study, researchers turned to machine learning to see if computers could learn the rules of hadron masses and then use those rules to guess the weights of particles it had never seen before. The team trained four different types of artificial intelligence models using data on 432 known, conventional hadrons. They fed the computers information about the quarks inside each particle and their quantum properties, asking the models to learn the relationship between these features and the particle's mass. Once the models were trained, the scientists tested them not just on new examples of ordinary particles, but on a completely separate group of 39 particles that were excluded from the training data. This group included exotic candidates and heavy baryons that had never been used to teach the models, effectively asking the computers to extrapolate: to guess the mass of something entirely new based only on what it had learned about the old.

The results showed that while all four models could predict the masses of ordinary particles with reasonable accuracy, they behaved very differently when faced with the unknown. One specific type of model, a neural network, proved to be far superior at guessing the masses of the unseen particles. When tested on 30 exotic or heavy states with known experimental masses, this neural network made errors that were roughly one-third the size of the errors made by the other three models. The other models, which rely on a different method of decision-making, struggled significantly. They tended to get stuck in a kind of logical trap where they could not distinguish between a particle with one heavy quark and a particle with two, leading to large mistakes. The neural network, however, managed to continue the trend it learned from the ordinary particles, correctly estimating that adding more heavy quarks would make the particle much heavier.

To understand why the neural network succeeded where the others failed, the researchers used a technique to look inside the "black box" of the computer's decision-making. They found that the model relied almost entirely on the presence of heavy quarks to determine the mass. When the model predicted the mass of a new particle, it looked at how many heavy quarks were present and adjusted the weight accordingly, a behavior that matched physical expectations even though the model had never seen those specific particles before. The study also used the neural network to predict the masses of nine doubly heavy baryons that have not yet been observed in experiments. While the predictions for these unobserved particles were slightly lower than some theoretical estimates, they provided a concrete benchmark for future experiments to check.

This work demonstrates that machine learning can be a powerful tool for exploring the frontiers of particle physics, provided the models are tested on data they have never seen. The study suggests that while computers can learn the underlying patterns of how quarks build mass, not all algorithms are equally good at applying those patterns to new situations. The success of the neural network in this specific task highlights that with the right approach, artificial intelligence can help physicists navigate the complex landscape of exotic matter, offering predictions that guide where to look next in the search for the universe's hidden building blocks.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →