This AI scientist can run experiments. Self-driving labs coming next

Northeastern University scientist Arun Bansil and his team have developed an AI scientist capable of autonomously running X-ray experiments and adapting its approach in real time, as reported in the July issue of Nature Machine Intelligence. The project, funded by the U.S. Department of Energy in collaboration with the SLAC National Accelerator Laboratory, marks a step toward self-driving labs where AI handles experimental tasks, potentially reducing the bottleneck of beam time at facilities like the Stanford Synchrotron Radiation Lightsource.

This AI scientist can run an X-ray experiment. Self-driving labs may come next. Northeastern researchers helped develop an AI scientist that can run X-ray experiments autonomously, adjusting when needed. The discovery paves the way for self-driving labs of the future. Forget self-driving cars. The “look ma, no hands” era of autonomy has gone beyond the road and into the lab. In the July issue of “Nature Machine Intelligence https://www.nature.com/articles/s42256-026-01261-5 ,” Northeastern University scientist Arun Bansil and his team present the AI X-ray scientist — an agentic model capable of setting up experiments and running them. Most importantly, it’s able to adapt its approach on the fly and change course when necessary. The development serves as the first step in the so-called self-driving labs of the future, where scientists let AI do the gruntwork as the human experts focus on the big questions. The project was funded by the U.S. Department of Energy DOE in collaboration with the Stanford Linear Accelerator Center SLAC National Accelerator Laboratory, a DOE research facility that houses equipment scientists use to study quantum matter, substances with features that can’t be explained with classical physics. Studying quantum materials can be done by~~ ~~shooting X-rays through a tiny single crystal of that matter and watching how they scatter, as a way to make conclusions about their atomic structure and electron dynamics. Traditionally, X-ray scattering happens in a huge circular particle accelerator known as the synchrotron. The one at SLAC is called the Stanford Synchrotron Radiation Lightsource SSRL and is about 768 feet in circumference — on par with a city block in size. Inside that facility is a large pipe bent into a circular shape. Scientists inject electrons into the pipe and accelerate them until they’re zooming around in circles, like runners on a giant racetrack. Naturally, electrons want to move in a straight line. Powerful magnets bend their path, forcing them to race around in circles. Every time electrons change direction, they give off energy in the form of X-rays. Scientists channel the X-rays from the circular pipe into straight tunnels called beamlines, with each tunnel serving as a miniature lab. They perform experiments at multiple beamlines at the same time, pointing the X-rays at the crystal they’re studying. In addition to SSRL, SLAC houses other light source facilities for studying materials and their properties. However, Bansil explained that many are oversubscribed — everyone wants to use them. Some only allow for one experiment at a time. As a result, the price tag climbs — quickly. X-ray experiments also call for complex preparatory work. “You just can’t walk in there and turn on a knob like you would in a kitchen,” Bansil said. In addition to setting up the equipment, adjusting and repositioning the crystal sample burns up a hefty chunk of precious beam time, he explained, adding that you end up spending a good part of it “just taking your sample and aligning it so that it can be useful to get data.” All in all, it’s “a real bottleneck,” he said. Enter the AI scientist. While based on a large language LLM model similar to the type of AI technology behind ChatGPT, the AI scientist can physically interact with its environment by moving objects, collecting data and reading detector outputs — things out of reach for a typical chatbot. Bansil said that the choice of words — scientist as opposed to agent — was deliberate. “A generic AI agent will simply execute a prescribed set of instructions,” he said. The AI scientist, on the other hand, reasons through the problem and adapts its approach based on observations — all without a handler micromanaging every step. “One gives it a broad guidance to start with, and then it pretty much does its own thing,” Bansil explained. The researchers put the AI scientist to the test in an actual SLAC X-ray facility by having it run through the steps of a simulated experiment. It passed with flying colors. The AI scientist sent commands to the diffractometer, analyzed the emerging pattern and made machinery adjustments when necessary. It figured out which tools to use for each step and pinned down the crystal’s orientation in space with the utmost precision, Bansil said. Editor’s Picks The moment of truth came when the AI scientist faced an unexpected snag involving a motor of a device used to orient the samples. Avoiding a potential fiasco, it spotted the glitch and regrouped on the fly, Bansil said. Better yet, it learned from the experience, applying the knowledge in the steps that followed. The ability to change course secured the AI scientist’s reputation as “a real scientific collaborator capable of being a highly valuable partner in the effort,” Bansil said. This project is part of a larger initiative at Northeastern University’s Quantum Materials and Sensing Institute QMSI where Bansil serves as the founding director. QMSI, which specializes in developing quantum materials and devices, is working on creating “the self-driving laboratories of tomorrow,” he said, and the AI Scientist marks the first step in making that vision a reality. Self-driving labs are not new. For example, Boston’s Atinary Technologies https://atinary.com/ , a self-driving chemistry lab located in Boston’s Seaport, crunches out an amount of weekly physical chemistry experimental data that’s on par with what a typical Ph.D. student produces over the course of their degree, according to Loïc Roch, who co-founded the lab together with Hermann Tribukait. Rigoberto Advincula https://www.ornl.gov/staff-profile/rigoberto-c-advincula , governor’s chair at the Department of Energy’s Oak Ridge National Laboratory, told Northeastern Global News that the bottleneck relief that happens as a result of combining the computational power of machine learning with the mechanical prowess of robotics has been nothing short of amazing. While his lab focuses on reaction chemistry, the science of creating new substances through chemical reactions, self-driving labs are also useful in the areas of drug discovery and material science, Advincula said. Bansil said that based on the results of the experiment, the future of self-driving labs looks more promising than ever. By intentionally using an off-the-shelf language model, his team proved that mainstreaming this technology won’t require training expensive custom-made AI — existing models can handle the complex workflows involved. The AI scientist hands researchers a win, enabling them to “spend less time adjusting knobs and waiting for hardware glitches, but instead focus on high-level experimental design, interpretation and creative scientific reasoning,” Bansil said.