In the laboratory, the path to discovery is often a long, winding road of trial and error. A physicist sets up a delicate experiment, adjusts a dial, waits for a signal, and then interprets the result. If the signal is weak or the equipment behaves strangely, the researcher must decide whether to tweak the settings, change the theory, or start over. This process relies heavily on human intuition, years of training, and the ability to spot when something is wrong. But what if a researcher had a tireless partner who could read every manual, run thousands of simulations in an afternoon, and never miss a tiny detail? This is the promise of artificial intelligence in science. However, giving a computer the keys to a physics lab is not as simple as asking it to solve a problem. Computers can make up facts, get lost in their own logic, or suggest dangerous actions if not guided carefully. The challenge is not just to build a smart machine, but to build a machine that thinks like a scientist: one that is honest about its mistakes, checks its own work, and knows when to ask for help.

A team of researchers at the University of Maryland has tackled this challenge by creating a new way for humans and artificial intelligence to work together. They did not simply program a computer to write code or run simulations; instead, they built a structured environment called PhysKit, designed to turn a large language model into a collaborative research assistant. Think of this system as a set of strict rules and helpful tools that guide the computer through the scientific method. The computer is not allowed to guess; it must follow a step-by-step process where it proposes an idea, tests it, checks the results for errors, and then reports back to the human researcher. The human remains the captain, setting the goals and making the final decisions, while the computer acts as a powerful engine that can explore possibilities far faster than a person could alone. This partnership allows the team to run "closed-loop" campaigns, where the computer works autonomously for hours or even days, iterating on solutions until it finds the best answer, all while keeping a perfect, organized record of every step it took.

The team tested this approach with six different experiments, ranging from designing new hardware to analyzing complex data. In one case, they asked the computer to improve a neural network, a type of software used to reconstruct images of electrical currents from magnetic measurements. The computer analyzed the existing design, tried thousands of variations, and discovered a new architecture that reduced errors by thirty-three percent. It did this by finding a more efficient way to process the data, using fewer computing resources while producing clearer results. In another experiment, the team wanted to understand how a specific defect in a two-dimensional material called hexagonal boron nitride behaves when hit with light. Scientists had been using a standard model to describe this behavior, but it did not match the data perfectly. The computer examined the data and proposed a new, more complex model with eleven energy levels instead of the previous seven or nine. This new model fit the experimental data much better and predicted how the material would behave under different conditions more accurately than any previous attempt.

The system also proved its worth in the physical world, not just in simulations. The researchers used the computer to design a microwave device needed to control quantum sensors. The computer had to balance many conflicting requirements, such as creating a specific type of wave while staying within the power limits of their equipment. It produced a design that was ready to be built and met all their strict experimental constraints. In a more direct test of hardware control, the computer helped commission a nuclear magnetic resonance spectrometer, a device used to detect magnetic fields. Starting from an uncalibrated state, the computer ran a series of experiments on the live machine, adjusting settings and analyzing the results until it achieved a frequency precision of about one hertz. This is a level of accuracy that would typically take a human expert days to achieve, but the computer did it in just over twenty hours. Finally, the team used the system to analyze data from an artificial spin-ice array, a material made of tiny magnetic islands. The computer successfully reconstructed the magnetic state of every single island in the array from blurry images, a task that had never been done before with this type of data.

What makes these results significant is not just that the computer found answers, but how it found them. The system was designed to be honest. If a proposed solution looked good on paper but failed a specific test, the computer rejected it. It kept a detailed log of every assumption it made and every piece of code it wrote, ensuring that a human could always trace back how a result was reached. The researchers found that the computer was particularly good at spotting when it was on the wrong track. For example, in the material study, the computer tested over a hundred different model structures. It found several that looked promising but discarded them because they failed to predict future data correctly or because they violated basic physical laws. It only selected the final model after it had passed every check, including tests where the computer had to predict data it had never seen before. This rigorous approach prevented the system from falling into the trap of simply memorizing the data or making up numbers to look successful.

The success of these experiments suggests a new way of doing science. Instead of replacing human researchers, this type of artificial intelligence acts as a force multiplier. It handles the repetitive, time-consuming work of testing ideas and checking calculations, freeing the human scientist to focus on the big picture and the difficult judgment calls. The computer can run experiments while the researcher sleeps, or it can explore a vast landscape of possibilities that would be impossible to navigate by hand. But it works best when there is a clear partnership. The human sets the boundaries, defines the goals, and provides the context that the computer cannot know on its own, such as the specific quirks of a piece of equipment or the subtle constraints of a physical system. The computer, in turn, brings a level of consistency and speed that transforms the research process.

The researchers emphasize that this is not a fully autonomous system that writes papers or makes discoveries on its own. It is a tool that requires a human to be in the loop, to verify the results, and to ensure that the science remains sound. The system is designed to be adaptable, learning from each project to improve its own framework. As the team continues to use PhysKit, they expect to add new skills and refine the process, making it even more effective for a wider range of scientific problems. The ultimate goal is to create a standard way for scientists to work with artificial intelligence, ensuring that the technology is used responsibly and effectively. By combining human expertise with the computational power of these agents, the team has shown that it is possible to accelerate discovery without sacrificing the rigor and honesty that are the foundation of science. The results are not just faster; they are better, leading to new models, better devices, and a deeper understanding of the physical world.