DOE-linked Synapse I project uses AI to deliver near–real-time imaging at national light sources
Loading...
Summary
Scientists described Synapse I, a foundation model in the DOE Genesis program that combines x-ray and neutron beamline data to invert detector images into object images in near real time, using large-scale GPU training and demonstrations at several DOE light-source facilities.
Synapse I, described by presenters at a Department of Energy demonstration, is a foundation-model effort to combine x-ray and neutron beamline data with machine learning so that experimenters can view processed images almost as soon as they are recorded. Presenter (S1) said the model aims to "accelerate the analysis over a 100 times," allowing real-time feedback and automated experiments.
The project couples coherent imaging methods with a large compute backbone, the presenters said. Presenter (S3) described a beamline demonstration run jointly by the Center for Nanoscale Materials and the Advanced Photon Source (APS), saying the APS upgrade increased usable flux at the beamline "by a factor of 20," turning experiments that previously took a day into experiments that now take about an hour.
Presenters emphasized how Synapse I alters experiment workflow. Before the model, researchers typically collected large data sets and returned to them later for analysis; Presenter (S2) said Synapse I delivers analyzed results to beamline scientists in real time so they can change or extend an experiment while it is running. "We can take an image as acquired on the detector and invert that into an image of the object that we just measured," Presenter (S2) said, describing the image-to-image mapping that enables immediate insight.
The team also addressed trust and validation. Presenter (S3) noted that coherent techniques can reveal features smaller than the x-ray beam and asked how users should trust AI-added details; Presenter (S1) said the beamline's complementary nanoscopy can map nanoscale objects directly and be used to validate the AI reconstructions. The presenters described training the foundation model on "over 1000 GPUs" at the Perlmutter cluster at Lawrence Berkeley National Laboratory and fine-tuning on Polaris at Argonne before deploying the system to APS beamlines.
Synapse I is part of the U.S. Department of Energy's Genesis project, the presenters said. Demonstrations have run at the Advanced Light Source at Lawrence Berkeley National Laboratory, the Advanced Photon Source at Argonne National Laboratory, and the National Synchrotron Light Source II at Brookhaven National Laboratory, and presenters said future demonstrations will expand to DOE light and neutron sources across the United States.
The demonstration positioned Synapse I as a tool for research on semiconductors, quantum devices and other advanced materials; presenters said the combination of upgraded light-source hardware and foundation-model software aims to make high-throughput, in situ experiments more accessible to chemists, materials scientists, physicists and biologists. The presenters did not announce procurement decisions, specific budgets, or deployment timelines beyond additional demonstrations.
