Northeastern University researchers have built an AI scientist that can manage complicated X-ray experiments without close human supervision. Published in Nature Machine Intelligence, the system can independently set up, carry out, and adjust to changes during experiments. This innovation highlights a big shift in how science is conducted, especially in advanced facilities like synchrotrons.
This AI scientist stands out by thinking and acting like a real researcher. It doesn’t just follow a script — it uses reasoning, analyzes data, and changes tactics based on what it sees. Arun Bansil, the lead scientist, emphasized that the AI works more like a partner than a tool, making its own choices and solving problems as they come up. 'A basic AI agent only follows a strict list of instructions,' Bansil said. The AI scientist, however, works through problems and adapts its methods based on what it observes — without needing someone to monitor each move.
X-ray Crystallography Explained
X-ray crystallography is vital in material science. This process involves sending high-energy X-rays through a crystal and watching how they scatter to learn about the material’s atomic structure and electron behavior. The equipment used is massive. At SLAC, the Stanford Synchrotron Radiation Lightsource is a 768-foot-long circular particle accelerator where electrons are accelerated in a loop by strong magnets. Inside, a large pipe is shaped into a circle. Scientists inject electrons into the pipe and speed them up until they orbit in circles, like runners on a track. Electrons naturally want to travel straight, but powerful magnets bend their path, keeping them in a loop. Each time the electrons change direction, they emit X-rays. These X-rays are guided from the curved pipe into straight tunnels called beamlines, which act like mini-labs.
Getting the crystal correctly aligned in the X-ray beam is a slow, manual process. Researchers often spend a lot of time adjusting the sample to get useful data, which eats into valuable beam time. The AI scientist, however, automates this process, making it faster and less work for human scientists. 'You can’t just walk in and flip a switch like in a kitchen,' Bansil noted. Adjusting and repositioning the crystal takes a big chunk of beam time, he added. 'You spend a good chunk of it just setting up the sample to get usable data.' Overall, it's 'a serious bottleneck,' he said.
The AI scientist was put to the test at a real X-ray facility at SLAC. It sent commands to the diffractometer, analyzed the scattered patterns, and adjusted when the crystal moved or the system had issues. During the test, it even noticed a faulty motor and fixed the problem without human help. When the AI scientist hit an unexpected issue with a sample-orienting motor, it quickly identified the problem and adapted. Bansil noted that the AI not only avoided a possible disaster, but also learned from the situation to improve its actions later.
What makes this AI valuable is its ability to learn and improve over time. After facing a problem, it uses the experience to make smarter choices in the future. This flexibility, according to Bansil, is key for the AI to act as a real research partner. Unlike traditional systems that stick to scripts, this AI improves through experience. The AI scientist is built on a large language model similar to the tech behind ChatGPT, but it interacts with its environment by moving objects, gathering data, and reading sensor outputs — all things a basic chatbot can’t do.
Implications for Self-Driving Labs
This project is a major step toward self-driving labs where AI handles repetitive and complex tasks, giving scientists more time for creative work. High-tech labs like those at SLAC are in huge demand, and the limited beam time is costly. The AI scientist has the potential to cut costs and streamline experiments. 'The costs go up fast,' Bansil said, noting that many facilities are overbooked — everyone wants to use them. Some can only handle one experiment at a time.
This work is supported by the U.S. Department of Energy and carried out in partnership with the Stanford Linear Accelerator Center (SLAC). SLAC operates several high-end facilities for studying materials, but many are constantly booked due to high demand. The AI scientist could help reduce these delays and make these tools more available and efficient.
Bansil made it clear that the AI isn’t meant to replace researchers, but to assist them. It takes over the routine parts of research so experts can focus on the most important and innovative work. 'You give it general guidance to start with, and then it mostly works on its own,' Bansil explained. The AI scientist’s success in the real-world test at SLAC 'confirmed its role as a real scientific collaborator and a highly valuable partner in the work,' Bansil said.