Lorentz-Equivariance without Limitations

In the high-energy world of particle physics, scientists smash protons together at nearly the speed of light to recreate the conditions of the early universe. The resulting debris is a chaotic spray of new particles, recorded by massive detectors as a cloud of points in space and time. To make sense of this chaos, researchers rely on the laws of special relativity, which state that the fundamental rules of physics do not change regardless of how an observer is moving or where they are located. This principle, known as Lorentz invariance, means that a particle's behavior looks the same whether viewed from a stationary lab or from a spaceship zooming past at high speed. However, when scientists use artificial intelligence to analyze these particle collisions, they face a difficult problem: standard computer programs do not naturally understand these rules of motion. They often treat the data as if it were a static picture, missing the deep geometric relationships that govern how particles interact. This limitation forces researchers to either build very specific, rigid computer models that are hard to adapt, or to use massive amounts of computing power to teach a standard model to guess the rules through trial and error.

A team of physicists and computer scientists has developed a new method called Lorentz Local Canonicalization, or LLoCa, which solves this problem by giving any neural network a built-in understanding of relativity without slowing it down. Instead of forcing the computer to learn the rules of motion from scratch, this method teaches the network to create a personal, moving reference frame for every single particle it sees. Imagine a particle as a traveler who carries its own map; the network predicts this map for each particle, allowing the computer to translate all the complex, moving data into a simple, local language where the laws of physics are easy to read. Once the data is translated into these local frames, the network can process it using any standard, flexible architecture. After the calculation is done, the network translates the results back into the global view of the detector. This approach allows the computer to respect the fundamental symmetries of the universe while remaining fast and adaptable to different types of physics problems.

The researchers tested this new framework on three distinct challenges that are central to modern particle physics: predicting the strength of particle interactions, generating realistic simulations of particle collisions, and identifying specific types of particles hidden within the debris. In the first test, they asked the network to predict the probability of a specific scattering event involving a Z boson and up to four gluons. They found that by adding this local reference frame system to standard neural networks, the predictions became significantly more accurate than those made by non-relativistic models, and they matched the performance of much more complex, specialized models. Crucially, the new method achieved this high accuracy with far less computational effort, running four times faster than the previous state-of-the-art models designed specifically for this task.

In the second application, the team used the method to generate new particle collision events from scratch. This is a vital task for simulating what detectors should see, but it is notoriously difficult because the computer must learn a complex, multi-dimensional landscape of possibilities. The researchers discovered that for this specific job, the network did not need to be perfectly symmetric in all directions. Instead, it performed best when it was allowed to break some of the strict symmetry rules to match the specific geometry of the particle detector, which has a preferred direction along the beam line. By letting the network learn to respect only the symmetries that actually exist in the detector data, they achieved the best possible results. This finding challenges the idea that a computer model must always be perfectly symmetric to be effective; sometimes, knowing where the symmetry breaks is just as important.

The final and most extensive test involved jet tagging, a process where computers must distinguish between jets of particles caused by heavy top quarks and those caused by ordinary background noise. This is a critical skill for finding new physics, as top quarks are often produced in rare, interesting events. The team applied their method to several established, high-performance network architectures, upgrading them to be relativistic. They created a massive new dataset containing 135 million simulated events to train these models, far larger than any previous dataset used for this specific task. On this large scale, the upgraded networks outperformed their non-relativistic counterparts and matched the performance of the most advanced specialized models. The results showed that while the new method works well on small datasets, its true power is revealed when the amount of data is large, allowing the network to leverage the laws of physics to learn more efficiently.

A key insight from the study concerns how the network handles the symmetry of the data. The researchers found that for some tasks, like generating events, it is sufficient for the network to respect only the residual symmetry of the detector. However, for identifying particles in jets, the best performance came from a hybrid approach. The network was allowed to use the full power of relativistic symmetry in its internal calculations, but it was also given specific reference points, such as the direction of the particle beam, as extra input. This allowed the network to decide for itself when to use the full symmetry and when to rely on the specific detector geometry. This flexibility proved to be the most effective strategy, allowing the model to achieve the highest accuracy without being constrained by rigid rules that might not apply to every part of the data.

The study also demonstrated that this new framework is not limited to one type of neural network. The researchers successfully applied it to both graph networks, which treat particles as connected nodes, and transformers, which are powerful models often used in language processing. In every case, the method acted as a universal upgrade, taking an existing network and making it relativistic with only a tiny increase in the number of adjustable parameters. The computational cost of this upgrade was minimal, adding only about one percent to the total number of parameters and requiring negligible extra time. This suggests that the method can be easily adopted by the wider physics community to improve existing tools without needing to rewrite entire software systems from scratch.

By providing a way to embed the fundamental symmetries of the universe into machine learning models with minimal overhead, this work removes a major barrier to progress in particle physics. It allows researchers to use the most powerful and flexible neural network architectures available today while ensuring they respect the laws of nature. The ability to generate better simulations, predict interaction strengths more accurately, and identify rare particles with higher confidence opens new doors for analyzing the vast amounts of data expected from future particle colliders. The researchers have made their code and datasets publicly available, inviting others to test and refine these tools, ensuring that the next generation of discoveries in particle physics can be built on a foundation of both data and deep physical understanding.