Meta的两款开源基础模型被部署到美国国家实验室，用于加速科学数据分析。其中Segment Anything Model 3（SAM 3）和DINOv3被应用于白宫Genesis Mission旗下的SYNAPS-I计划中，该计划旨在利用人工智能推进科学发现[1]。
这些模型被部署在劳伦斯伯克利、阿贡、布鲁克海文、橡树岭和SLAC五家国家实验室，专门用于分析X射线和中子科学实验数据[1]。部署规模达300个A100 GPU，跨越五家实验室的60名研究人员参与其中[1]。
在实际应用中，这些模型大幅提升了数据处理效率。针对图像分割任务，原本需要专家耗费一个月进行的手工标注工作，现在仅需约15分钟即可完成[1]。这使得科学家能够实时分析实验数据，而非等待离线处理结果。
美国能源部面临的数据规模不断扩大。能源部每年产生数十PB的数据量，相当于200万小时的高清视频[1]。与此同时，科研设备的检测能力也在快速升级，从原来每6秒捕获一张图像的频率提升到每秒100,000张[1]。
应用已见实效。在葡萄树干旱耐受性研究中，研究人员使用微CT扫描配合这些模型识别木质部导管，展现了AI模型在生物科学领域的潜力[1]。
美国能源部副部长Dario Gil对此表示："这将发现时间从几天压缩到瞬间，建立持续自我改进的科学模型"[1]。
Meta's open-source foundation models are accelerating scientific discovery at U.S. national laboratories as part of the White House Genesis Mission initiative. The Segment Anything Model 3 (SAM 3) and DINOv3 have been deployed within the SYNAPS-I program to analyze X-ray and neutron science data, with the goal of using artificial intelligence to expedite research breakthroughs. [1]
The deployment spans five national laboratories—Lawrence Berkeley, Argonne, Brookhaven, Oak Ridge, and SLAC—involving 60 researchers across these institutions. [1] The models are running on 300 A100 GPUs and are being applied to image segmentation tasks that previously consumed substantial time and expert resources. [1] What once required approximately one month of manual annotation by specialists can now be completed in roughly 15 minutes, enabling scientists to analyze experimental data in near real-time. [1]
The scale of data being processed is enormous. The Department of Energy generates tens of petabytes of data annually—equivalent to 200 million hours of high-definition video—from its scientific instruments. [1] Recent detector upgrades have intensified this challenge: equipment that previously captured one image every six seconds now captures 100,000 images per second. [1]
The practical impact is already evident. One application involves studying drought tolerance in grapevines through micro-CT scanning, where the AI models identify xylem vessels in wood samples with greater efficiency. [1] According to Deputy Secretary of Energy Dario Gil, the deployment will "compress discovery time from days to moments, establishing continuously self-improving scientific models." [1]