Meta has disclosed that its open-source vision models SAM 3 and DINOv3 are powering a new pipeline called SYNAPS-I, a project run by Lawrence Berkeley National Laboratory and a consortium of DOE national labs. The pipeline is designed to automate the segmentation of scientific images at facilities that study materials using powerful light and neutron sources.

SYNAPS-I (SYnergistic Neutron And Photon Science – Intelligence) is a Genesis Mission project run in partnership with Argonne, Brookhaven, Oak Ridge, and other national laboratories. Its target is a bottleneck that has slowed facility science for years: manual segmentation of experimental images, which consumes weeks of expert effort per dataset. Segmentation, in this context, means identifying and outlining individual structures within an image so they can be measured and analyzed. The pipeline applies SAM 3 and DINOv3 to replace or augment that manual process with automated, pixel-level segmentation.

The two models fill complementary roles. DINOv3 is a self-supervised vision model, meaning it learns visual patterns from raw images without requiring human-labeled training data. SAM 3 draws precise pixel-level boundaries around individual objects in an image, performing the function that a scientist would otherwise do by hand. Together, they form a pipeline that can operate on the voluminous image data streaming out of upgraded detectors.

Those detectors are the reason the bottleneck exists in the first place. DOE light and neutron source facilities have upgraded from capturing one image every six seconds to 100,000 images per second, increasing data generation by orders of magnitude compared to a decade ago. The facilities now produce tens of petabytes of data annually, roughly equivalent to 2 million hours of HD video. Lawrence Berkeley's Advanced Light Source, a football field-sized facility producing intensely bright X-ray beams for studying materials at atomic and molecular scales, is one such site. The data volume has long since exceeded what manual analysis can handle, which is precisely the problem SYNAPS-I was designed to address.

The Genesis Mission, launched by the White House in late 2025 and led by the DOE, is a national initiative to accelerate scientific discovery using advanced AI. The DOE announced funding to advance the mission's efforts in March 2026, uniting national labs, industry, and academia to apply AI to problems in energy, discovery science, and national security. The mission's scope is broad: it encompasses "self-driving" labs that integrate AI with robotics and real-time analysis to automate scientific workflows from hypothesis generation onward, AI-driven transformation of the building lifecycle from design through operation, and identified needs in edge computing and AI accelerators where bottlenecks persist in design, density, scaling, fabrication, and integration. SYNAPS-I is one project within that larger program.

The choice of open-source models for a national science pipeline is worth noting. Meta has released SAM 3 and DINOv3 as open-source foundation models, meaning the code and model weights are available for external researchers to inspect, modify, and deploy. For a multi-lab consortium operating under federal science missions, that openness matters: it allows reproducibility, on-premises deployment without vendor API dependencies, and the ability to fine-tune models on domain-specific scientific imagery without licensing constraints. Whether that openness translates into sustained adoption across the DOE complex will depend on how the models perform under the specific imaging modalities these facilities produce, which differ substantially from the natural-image distributions these models were trained on.

The historical parallel is hard to ignore. In the 2000s and 2010s, the rise of high-throughput DNA sequencing created a data deluge in genomics that required an entire bioinformatics ecosystem to absorb. The DOE's detector upgrades have created a structurally similar problem in materials science and photon science. The question is whether foundation models pretrained on general visual data can generalize effectively to X-ray diffraction patterns, neutron scattering images, and other specialized modalities. Meta's blog post reports that SYNAPS-I is actively using the models in its pipeline, which suggests at least functional performance, though the post does not detail quantitative benchmarks against expert manual segmentation.

What SYNAPS-I enables, if the pipeline holds up, is straightforward: faster turnaround from experiment to scientific insight at facilities that cost hundreds of millions of dollars to operate. Every week a researcher spends manually tracing boundaries in images is a week of beam time not being analyzed, a finding not being reported, and a next experiment not being designed. The technical bet is that self-supervised pretraining plus promptable segmentation can match or approach expert-level performance on scientific images at a scale no team of human analysts could sustain. That bet is now being tested in production at one of the DOE's premier light sources.

The broader context here is one I have watched play out before. When my older child started working in a university imaging lab a few years ago, she spent the better part of a summer semester hand-tracing cell boundaries in microscope images. The work was valuable but slow, and it was exactly the kind of task that a well-trained segmentation model could accelerate without replacing the scientific judgment behind it. If SYNAPS-I delivers on its premise, the researchers at these facilities get to spend their time on the science rather than on the outlining, and that is a trade worth making.