The Lawrence Berkeley National Laboratory, a premier U.S. Department of Energy research facility, faces a data deluge from its Advanced Light Source, generating tens of petabytes annually. Upgraded detectors now capture 100,000 images per second, overwhelming traditional manual analysis methods. The volume and complexity of data have created a bottleneck in scientific imaging, where segmentation — the process of identifying and labeling structures within images — is critical for research.
Meta's Segment Anything Model 3 (SAM 3) and DINOv3 are open-source foundation models designed to address this challenge. DINOv3, a self-supervised vision model, understands visual patterns without human-labeled data, while SAM 3 draws precise boundaries around objects in images. Together, they form a pipeline that transforms raw X-ray data into semantically labeled 3D volumes, enabling real-time interpretation during experiments. The models were fine-tuned on scientific imaging data from DOE beamlines and deployed across 300 A100 GPUs at national supercomputing facilities, achieving a 15-minute turnaround for data analysis.
The SYNAPS-I team demonstrated the pipeline's effectiveness by analyzing micro-CT scans of grapevines to study drought resilience at the cellular level. Previously requiring a month of expert annotation per time step, the process now takes just 15 minutes, allowing scientists to study dynamic biological processes in real time. This open-source approach enables national labs to manage data securely and adapt models for scientific domains, fostering collaboration across facilities.
Source: metaai
