FUSION: a skill-based research agent for publicly obtainable nuclear-physics codes

In the world of nuclear physics, scientists study the tiny, dense cores at the heart of atoms to understand how matter holds together and how stars burn. To do this, they rely on complex computer programs that act as virtual laboratories. These programs take a set of instructions describing a specific atomic collision or reaction and calculate the likely outcome. For decades, running these simulations has been a delicate task. A researcher must find the right software, often hidden on an old personal website, install it on a computer that might not be compatible, and then write a file of instructions that follows very strict, often obscure rules. If the rules are followed perfectly, the computer prints a result. But if the rules are followed with a subtle mistake—like using the wrong unit for a distance or the wrong definition for a radius—the computer will still print a smooth, plausible-looking number. The danger is that the result looks correct even though the physics behind it is wrong.

This is the problem that Jin Lei, a physicist at Tongji University, set out to solve with a new tool called FUSION. The name stands for a skill-based research agent designed to handle publicly available nuclear-physics codes. Instead of asking a computer to guess how to write a simulation from scratch, FUSION acts as a rigorous assistant that follows a strict checklist. It does not just run the code; it verifies that the code is running the right physics. The system works by taking a known, correct example of a calculation and modifying it slightly to fit the user's new question, rather than inventing a new set of instructions from memory. Before it ever reports a result to a human, it runs a series of tests to ensure the output matches a known benchmark to a high degree of precision. If the numbers do not line up exactly as they should, the system stops and flags the error, preventing a silent mistake from being published as a discovery.

The tool is built like a layered system to keep the user's private research notes separate from the public software. At its core, it contains twenty different "skills," each one dedicated to a specific nuclear-physics program. These skills are not just simple scripts; they are complete packages that know how to download the original source code from the author, build it on the user's computer, and run it in a clean environment. Each skill includes a library of verified examples that have already been proven to work. When a user asks for a calculation, the agent selects the closest verified example and adjusts it, ensuring that the starting point is always solid. It also includes a massive, offline collection of over 61,000 scientific papers from the last thirty years. This library allows the agent to search for context and methods without needing an internet connection, but it is designed to be a reference tool, not a source of truth; the agent is instructed to always point back to the original paper rather than relying on its own summary.

The creators of FUSION tested the system rigorously before releasing it. They did not just check if the software could run; they tried to break it. They used a second, different computer program to act as a critic, reading the instructions and running the code to find hidden flaws. In one instance, this process found nineteen errors in a single skill, and further rounds of testing uncovered more mistakes that were accidentally introduced while fixing the first ones. This adversarial testing ensured that the system could catch subtle failures, such as a program that reports a successful run even when it has crashed, or a script that fails only when the computer's language settings are changed. The result is a collection of twenty verified skills covering a wide range of nuclear physics, from how atomic nuclei scatter off one another to how stars explode and how heavy ions collide.

To demonstrate how the system works in a real scenario, the author ran a calculation for the scattering of a neutron off a zirconium nucleus at a specific energy. The agent loaded the correct instructions for the FRESCO program, a standard tool for this type of physics. It noticed a potential trap: the program and the physical model it was using defined the size of the nucleus in slightly different ways. If the agent had simply written the instructions from memory, it would have used the wrong definition, making the nucleus appear 22 percent larger than it should be. This mistake would have produced a result that looked reasonable but was physically incorrect. Because FUSION started from a verified example and checked the math, it caught the discrepancy and corrected the instruction. The system then ran the calculation and compared the result against a second, completely different type of computer program. The two programs agreed on the answer to within a tiny fraction of a percent, confirming that the calculation was being performed correctly. Finally, the agent checked a database of real-world measurements and found that no experiment had ever been done at that exact energy, so it reported the absence of data and instead compared the calculation to measurements taken at nearby energies, showing how well the theory matched reality.

The FUSION system is now available for anyone to use, provided they have a computer capable of running the software. It does not include the physics programs themselves, which remain under the copyright of their original authors, but it provides the bridge that allows them to run safely and correctly. The project represents a shift in how scientific software is managed, moving away from informal, memory-based knowledge toward a system of verified, reproducible steps. By ensuring that the computer is not just running, but running the right physics, FUSION helps scientists avoid the quiet errors that can derail years of research. It does not replace the judgment of the physicist, who must still decide if the model fits the problem, but it removes the risk of the tool itself failing in a way that looks like success. In a field where a single misplaced decimal can change the meaning of a discovery, this kind of verification offers a new level of confidence in the numbers that drive our understanding of the atomic world.