Infrared Subtraction with Artificial Intelligence

In the heart of modern physics, researchers study the fundamental building blocks of the universe by smashing particles together at incredible speeds. When these collisions occur, they do not just produce a few simple fragments; they often spawn a shower of new particles that fly out in every direction. To understand the laws governing these events, scientists must calculate the probability of every possible outcome with extreme precision. However, a major hurdle exists in these calculations: when a particle emits a burst of energy that is too faint to be seen or when two particles move in almost the exact same direction, the mathematical formulas used to describe the event break down and produce infinite, nonsensical numbers. These are known as infrared singularities. While the laws of physics guarantee that these infinities cancel each other out in the final result, separating them from the meaningful data to get a precise number for a specific experiment has been a decades-long struggle for theoretical physicists.

A team of researchers at Beijing Normal University and the Chinese Academy of Sciences has now tackled this problem using a new approach that blends human physical insight with the creative problem-solving abilities of artificial intelligence. They developed a method to isolate and remove these troublesome infinities without relying on the traditional, rigid mathematical shortcuts that have been used for years. Instead of forcing the calculation to fit a pre-existing mold, they used a large language model—a type of artificial intelligence trained on vast amounts of text and code—to design new ways of mapping the chaotic spray of particles back to their simplest starting point. The researchers found that by letting the AI devise these maps and then verifying them against established theories of particle behavior, they could reconstruct complex particle collision data with high accuracy. This work demonstrates that artificial intelligence can not only speed up calculations but can also help construct the very mathematical tools needed to solve some of the most difficult problems in high-energy physics.

The core challenge the team addressed is how to handle the "noise" of particle collisions. When particles collide, they emit radiation, much like a car engine emits exhaust. In the quantum world, this exhaust can be so faint that it is impossible to distinguish from the background, or so aligned with other particles that they appear as a single stream. Standard calculations struggle with these fuzzy boundaries. For years, physicists have used a technique called "slicing," which involves drawing an arbitrary line in the data to separate the clear, measurable events from the fuzzy, problematic ones. While this works, it introduces its own errors and requires massive computing power to make the line thin enough to be accurate. The researchers wanted a better way: a method that could handle the entire range of particle emissions at once, without drawing an arbitrary line, while still canceling out the mathematical infinities.

To achieve this, the team built a framework that separates the calculation into two parts. The first part deals with the radiation that is actually emitted, and the second part deals with the "contact" term, which represents the finite, clean contribution of the collision at its simplest level. The key innovation was using an artificial intelligence agent to design the rules for how to project the complex, multi-particle final state back onto the simple, two-particle starting state. In one version of their experiment, the AI used a neural network—a type of computer program modeled after the human brain—to learn how to map the chaotic particle spray to a clean starting point. In another version, the AI generated a strict, step-by-step mathematical recipe to do the same thing. In both cases, the human researchers provided the physical laws and constraints, telling the AI what was possible and what was forbidden, while the AI figured out the specific steps to get there.

The team tested their new method by simulating electron-positron collisions, a standard process in particle physics where an electron and its antimatter counterpart annihilate to produce jets of particles. They focused on scenarios where three or four jets of particles were produced, which are significantly more complex than the simpler two-jet cases. The results were striking. The AI-designed methods successfully reconstructed the full theoretical predictions for these collisions, matching the results of the most sophisticated existing simulation tools with high precision. The researchers found that the AI could determine the necessary "contact" terms—the mathematical corrections needed to cancel out the infinities—either by fitting the data to known theoretical patterns or by calculating them directly using a new, local formula that did not require the arbitrary slicing line. This direct calculation allowed them to reuse existing lower-order calculations and combine them with advanced theoretical predictions to get the final answer.

Perhaps most surprisingly, the team achieved these results using very modest computing resources. The entire numerical calculation and the training of the neural network were performed on the central processing unit of a 2020 Apple MacBook, a standard laptop, without any specialized graphics cards or supercomputers. This suggests that the new method is not only accurate but also efficient enough to be run on local machines, potentially democratizing access to high-precision physics calculations. The researchers also extended their work to even more complex scenarios, including a two-jet collision at a higher level of precision, where they used machine learning to reduce the statistical noise in the results. By training the AI to recognize patterns in the data and subtract them out, they stabilized the calculation, making the final numbers more reliable.

The implications of this work go beyond just getting a better number for a specific experiment. It represents a shift in how theoretical physics is done. For decades, the construction of the mathematical tools needed to cancel out infinities has required immense human effort and deep analytical insight, often taking years to develop for a new type of collision. This study shows that artificial intelligence can assist in the creative process of building these tools. The AI did not just crunch numbers; it devised the structure of the subtraction method itself, proposing new ways to map particles and calculate corrections that human researchers might not have considered. The researchers verified that their AI-generated maps respected all the necessary physical laws, such as the conservation of energy and momentum, and that they correctly handled the tricky limits where particles become indistinguishable.

While the study is a proof of concept, it opens a clear path forward. The researchers demonstrated that their approach works for three and four jets at a standard level of precision and for two jets at an even higher level. They also outlined how the method could be extended to even more complex collisions involving three jets at that higher precision level, providing the necessary mathematical maps and formulas for future work. The success of the method relies on a partnership between human intuition and machine learning. Humans define the physical boundaries and the goals, ensuring the AI stays within the laws of nature, while the AI explores the vast space of possible mathematical solutions to find the most efficient path. This collaboration allows for the reuse of existing calculations and the integration of advanced theoretical predictions, creating a more robust and flexible framework for future discoveries.

The findings suggest that the future of precision physics may involve a new kind of workflow where artificial intelligence acts as a co-pilot in the construction of theoretical tools. By automating the design of subtraction formulas and improving the numerical integration of complex data, AI can help physicists push the boundaries of what can be calculated. The ability to run these sophisticated calculations on a laptop indicates that the barrier to entry for high-precision work could be lowered, allowing more researchers to tackle difficult problems. As the field moves toward even higher levels of precision, where the calculations become exponentially more difficult, the ability to construct these subtraction methods automatically could become essential. The researchers have shown that with the right guidance, artificial intelligence can not only solve equations but can also help invent the methods needed to solve them, turning a decades-old bottleneck into a manageable step in the journey of discovery.