The U.S. Department of Energy's Argonne National Laboratory has demonstrated a new capability of agentic artificial intelligence. The system allows scientists to control complex experiments using ordinary conversational language.
The demonstration took place within the SYNAPS-I project — Synergistic Neutron and Photon Science – Intelligence. The project brings together data from neutron, X-ray, and microscopy experiments from several national laboratories and is part of the Genesis Mission initiative, led by Lawrence Berkeley National Laboratory.
In the first phase of SYNAPS-I, researchers created the PtychoFM model for fast real-time reconstruction of ptychographic images. Now an agentic AI layer has been added, which moves from data processing to controlling the experiment itself.
Scientists can simply say: "measure this area" or "produce a high-resolution map." The system gathers data on its own, analyzes images, makes decisions, and, when necessary, requests human confirmation for risky actions.
In the demonstration, the agentic AI successfully located the desired region on a microelectronics sample at the 26-ID Hard X-ray Nanoprobe beamline, jointly operated by APS and CNM. After detecting the region, the system automatically switched to high resolution.
Additionally, the agent is connected to the Meta Segment Anything Model 3 (SAM3) for segmenting and labeling objects in images. This makes the analysis even more accurate and faster than traditional methods.
"Instead of writing code or entering technical commands, researchers can simply talk to the AI, which automatically performs data collection, processing, and analysis," noted Argonne computational scientist Ming Du.
The project combines the efforts of Argonne, LBNL, Brookhaven, SLAC, and Oak Ridge. The demonstration took place on 5 October 2026 and became an important step toward fully autonomous laboratories.
How far can such autonomy go in the coming years?
According to Matthew Cherukara, head of Argonne's SYNAPS-I team, the combination of fast reconstruction, agentic control, and segmentation lays the foundation for self-driving microscopes and closed-loop experiments in a wide range of fields — from semiconductors to quantum devices.
