Meta Vision Models Cut Scientific Image Analysis to 15 Minutes

Lawrence Berkeley National Laboratory operates the Advanced Light Source (ALS), a facility that uses exceptionally bright X-ray beams to examine matter across scales ranging from atoms and molecules to entire plant structures. Its beamline instruments have become far faster and more precise, but those improvements have created a major data-analysis problem. Detectors that once captured one image every six seconds can now record as many as 100,000 images per second. Across the US Department of Energy’s light and neutron facilities, instruments consequently produce tens of petabytes of data each year—millions of gigabytes, comparable to roughly two million hours of high-definition video.

Scientists cannot process this flood through traditional manual workflows. The shortage of specialists makes the bottleneck worse, because interpreting scientific images requires substantial domain knowledge. The challenge is particularly serious for in-situ experiments, which follow events such as chemical reactions, biological responses, or material failures while they happen. These experiments are most valuable when researchers can interpret results immediately and adjust the work in progress, yet human teams cannot manually analyze images at anything close to the rate at which modern instruments generate them.

A central task in this workflow is image segmentation: identifying meaningful structures and tracing their exact boundaries. Segmentation turns an undifferentiated field of grayscale pixels into a map whose components—such as cell walls, mineral grains, plant vessels, or semiconductor layers—can be measured and compared. The same general computer-vision technique helps distinguish tumors from healthy tissue in medical scans and pedestrians from roads in autonomous-driving systems. In laboratory research, however, experts have traditionally spent days or weeks annotating each dataset, making segmentation one of the most acute obstacles to rapid scientific interpretation.

The White House launched the Department of Energy-led Genesis Mission in late 2025 to use advanced AI to accelerate scientific discovery and strengthen technological leadership. One of its early efforts is SYNAPS-I, short for Synergistic Neutron And Photon Science–Intelligence. Developed with Lawrence Berkeley, Argonne, Brookhaven, Oak Ridge, and other national laboratories, the project seeks to replace monthslong X-ray and neutron data-analysis bottlenecks with a system capable of producing useful results while an experiment is still underway. Scientific imaging, and segmentation in particular, is a major focus.

The SYNAPS-I pipeline combines two open-source Meta foundation models—Segment Anything Model 3 and DINOv3—to turn raw experimental images into semantically labeled three-dimensional reconstructions. DINOv3 is a self-supervised vision model, meaning it learns recurring visual patterns from unlabeled images rather than depending entirely on expensive human-prepared examples. It supplies broader context about the structures present in an image and their locations. SAM 3 complements that capability by drawing accurate boundaries around individual objects at the pixel level, performing in seconds a task that a scientist might otherwise complete carefully by hand.

The research team adapted both models using specialized scientific images gathered at DOE beamlines. It then deployed the pipeline across 300 Nvidia A100 GPUs at national high-performance computing facilities including NERSC. SAM 3 provides detailed object outlines, while DINOv3 helps determine what the structures represent and how they relate to the surrounding sample. After processing and reconstruction, the system sends a labeled 3D volume back to the researcher standing at the beamline. The complete turnaround is approximately 15 minutes, allowing interpretation before the experiment has finished rather than weeks afterward.

Researchers demonstrated the approach by investigating how grapevines respond to drought at the cellular level. Micro-CT scans recorded at the Advanced Light Source were reconstructed as 3D volumes of vine stems. The AI pipeline automatically identified xylem vessels, the microscopic tubes that carry water through a plant. Tracking changes in these vessels as drought develops can reveal how water transport breaks down or adapts under stress, potentially supporting future research into more drought-resilient crops and broader agricultural resilience.

Previously, annotating one time step in this kind of experiment could require about a month of expert work. SYNAPS-I reduced that interval to around 15 minutes. The significance is not merely cheaper image processing: scientists can now examine changing biological processes at approximately the pace at which instruments collect the underlying data. The project illustrates how general-purpose, open-source vision models can become practical scientific infrastructure when fine-tuned on domain-specific data and connected to national laboratory supercomputers. It also shows the intended role of the Genesis Mission’s first projects: using existing AI advances to remove concrete constraints on experimentation and convert rapidly growing instrument output into information researchers can act on in near real time.