The bottleneck in advanced science facilities is no longer only the experiment. It is the analysis that follows. Meta’s account of Berkeley Lab’s early Genesis Mission work puts concrete numbers on the problem: U.S. Department of Energy light and neutron source facilities now generate tens of petabytes of data each year, while upgraded detectors have moved from one image every six seconds to as many as 100,000 images per second. Manual analysis cannot keep up with that rate.

The project is called SYNAPS-I. Led by Lawrence Berkeley National Laboratory, it brings together 60 researchers across five national labs. The goal is to turn beamline facilities such as the Advanced Light Source into intelligent discovery platforms. Instead of collecting data first and analyzing it much later, the system would segment images, recognize structures, help generate hypotheses, recommend next experiments, and transfer knowledge across facilities. Meta says its SAM 3 and DINOv3 open-source vision models are part of that first wave of projects.

The Advanced Light Source is a football-field-sized facility that uses intense X-ray beams to study materials from atomic and molecular scales up to biological samples. As instruments improve, the data volume rises by orders of magnitude. The immediate value of AI is not abstract productivity; it is reducing the backlog between measurement and decision. If models can identify relevant regions, track patterns, and surface anomalies while data is being produced, scientists can adjust an experiment before beam time is gone.

For Meta, this is a strategic proof point for open-source vision models in national-priority science. For labs, it is a move toward foundation models inside the experimental loop rather than just downstream analysis. The caveat is important: the article describes early Genesis Mission projects and their intended architecture, not a finished autonomous science platform. The next evidence to watch is operational validation: how SAM 3 and DINOv3 perform across different instruments, how error bounds are handled during real-time decisions, and whether knowledge learned at one beamline improves experiments at another.

#sam-3

Meta models move into a 100,000-images-per-second science bottleneck