SMEFT-Pheno-Agent: a natural-language-driven AI agent for machine-learning-assisted Standard Model Effective Field Theory phenomenology

Imagine the universe as a giant, cosmic puzzle. Physicists have spent decades assembling most of the pieces, creating a picture called the "Standard Model" that explains how tiny particles like electrons and quarks interact. It's a brilliant map, but we know it's incomplete. There are gaps—mysterious forces and heavy particles we can't see yet. To find them, scientists smash particles together at mind-boggling speeds in giant machines called colliders. They look for tiny cracks in the map, hints that the rules of physics might be slightly different than we thought. This is where "Effective Field Theory" comes in. Think of it as a flexible ruler: instead of guessing exactly what the missing piece looks like, scientists measure how much the ruler bends under different conditions. If the bend is weird, it suggests something new is hiding there. The problem? Doing these measurements is like trying to bake a cake while simultaneously building the oven, writing the recipe, and counting every single crumb. It requires a massive chain of complex computer programs, and if one step is done wrong, the whole cake collapses.

Enter SMEFT-Pheno-Agent, a new digital assistant designed to be the ultimate sous-chef for these physics experiments. This isn't a robot that invents new laws of physics; rather, it's a highly organized project manager that speaks human language. Imagine you tell a very smart, very strict robot, "I want to see what happens if we smash muons together at 10 TeV and look for a specific kind of bend in the rules." The robot doesn't just nod; it takes that sentence and automatically builds the entire kitchen. It checks if the ovens (computer programs) are working, sets the temperature (physics parameters), orders the ingredients (particle events), and even decides which tools to use for the final taste test (machine learning).

The paper introduces this agent as a workflow that turns a simple conversation into a rigorous, 12-step scientific study. The authors, Yu-Chen Guo, Jie Wang, and Ji-Chong Yang, built a system where an AI agent acts as the conductor of an orchestra. The agent listens to your request in plain English, then translates it into a locked set of instructions that specialized physics software (like MadGraph5_aMC@NLO and Delphes) can follow. The magic here is that the AI never touches the actual math or the numbers. It's like a chef who writes the recipe and tells the sous-chefs exactly what to do, but never actually stirs the pot or tastes the soup. This ensures that the "cooking" is done by trusted, proven tools, while the AI handles the messy job of keeping everything in order.

The system works in twelve distinct phases, moving from a simple chat to a final, audited report.

1. The Intake: You talk to the agent, specifying the collision type, energy, and what you want to find. The agent locks these details into a configuration file so nothing can accidentally change later.
2. The Setup: It checks that all the necessary computer programs are installed and ready to go.
3. The Simulation: It runs the heavy lifting: generating millions of simulated particle collisions, simulating how detectors would see them, and creating "pure" examples of what new physics might look like.
4. The Machine Learning: This is where the agent gets clever. It looks at the data and suggests the best mathematical tools (algorithms) to spot the difference between "normal" physics and "new" physics. It picks a specific "operating point"—a threshold for what counts as a hit—and sticks to it for the whole study.
5. The Analysis: It scans through different possibilities, fits the data to a curve, and calculates how strong the evidence is for or against the new physics.
6. The Audit: Finally, the agent checks its own work. It verifies that every file exists, every number matches the plan, and the story makes sense. It even drafts a manuscript for scientists to read, citing exactly where every number came from.

The paper demonstrates this with a specific example: simulating a collision at a future muon collider with an energy of 10 TeV. The agent successfully guided the process, generating event samples, training machine learning models (like AdaBoost), and producing a statistical interval for a specific coefficient ($c_{gT,0}$). The results showed that with a specific threshold of 0.78, the system could constrain the physics parameter to a range of roughly $[-0.948, 1.531] \times 10^{-3} \text{ TeV}^{-4}$ at a significance level of 2.

Crucially, the authors are careful to state what this tool is not. It is not a magic wand that solves all physics problems, nor does it replace the need for human scientists to understand the theory. The agent cannot invent new physics, and it doesn't claim to have found a new particle. It is a tool for reproducibility. In the past, if a scientist wanted to repeat a complex study, they might have to spend weeks re-figuring out how the previous person set up their computer programs. With SMEFT-Pheno-Agent, the entire process is recorded in a "manifest"—a digital receipt that proves exactly what was done, how it was done, and what the results were. If you want to replay the study, the agent can do it again, exactly the same way, because it never relies on guesswork or memory.

The paper argues that this approach solves a major bottleneck in modern physics: the "handoff" problem. Usually, moving from one computer program to another is a fragile, manual process where errors can sneak in silently. By automating these handoffs and keeping a strict, unchangeable log of every step, the agent ensures that the science is solid. The authors suggest that while this specific version focuses on one type of physics scenario, the framework could eventually be expanded to handle more complex, multi-parameter studies. For now, it stands as a proof-of-concept: a way to let AI handle the paperwork and the plumbing, so physicists can focus on the big questions.