AI & Scientific Computing

How Open-Source AI Cut Synchrotron Analysis From Months to 15 Minutes

Meta’s SAM 3 and DINOv3 vision models, retrained on X-ray and neutron data, now deliver a labeled 3D reconstruction to a scientist at the instrument within the time it takes to finish a coffee.

• The pipeline uses two foundation models working in tandem: DINOv3 reads the whole image for context, then SAM 3 draws pixel-precise boundaries around each structure.
• Fine-tuned on data from DOE beamlines and run on roughly 300 A100 GPUs at facilities like NERSC, it returns a semantically labeled 3D volume in about 15 minutes.
• A demonstration on grapevines under drought tracked the microscopic water-carrying xylem vessels as they changed over time — a task that once took a month of manual annotation.

The bottleneck at DOE’s light and neutron sources is no longer the beam — it is the backseat. Modern detectors at the Advanced Light Source and its partner facilities can capture 100,000 images per second, up from one every six seconds a decade ago. That surge now produces tens of petabytes of data a year, and the specialist who can turn a grayscale X-ray slice into a map of cell walls, mineral grains, or semiconductor layers is simply overmatched.

The breakthrough is in segmentation — the computer-vision task of tracing clean boundaries around distinct objects in an image. It is the same primitive that separates a tumor from healthy tissue in a scan or a pedestrian from a road in a self-driving car. SYNAPS-I, a multi-lab platform born of the national Genesis Mission, dropped Meta’s open-source DINOv3 and SAM 3 into the synchrotron workflow, fine-tuned them on real beamline imagery, and wired the result back to the operator in near real time. What used to be a months-long hand-labelling chore is now a 15-minute delivery.

Two details are quietly important. First, the work runs on government infrastructure, not external cloud — an open-source foundation model is essential because national labs must keep prepublication data inside secure boundaries. Second, the grapevine test shows the scientific payoff: watching water transport collapse in a stem under drought could seed genuinely drought-resilient crops. The model choice is therefore not a cosmetic swap; it is the lever that makes the entire measurement loop possible at the speed of data itself.