Real-Time AI Transforming Large-Scale Imaging at National Labs

Photo: Beamline scientist Tao Zhou explains the capabilities of the 26-ID beamline, where the SYNAPS-I AI platform converts X-ray diffraction data into high-resolution images in real time, dramatically accelerating analysis at US Department of Energy facilities. CREDIT: MARK LOPEZ/ARGONNE NATIONAL LABORATORY

As advanced discovery tools flood US Department of Energy (DOE) laboratories with data, scientists face a growing problem: humans can’t keep up with the volume and speed of information being generated.

Modern X-ray, microscopy, and neutron facilities produce vast streams of high-value imagery, but tools to interpret this data at scale haven’t kept pace. Researchers are increasingly turning to artificial intelligence (AI) to convert these massive datasets into usable insight.

As part of DOE’s Genesis Mission—a national effort to transform American science and innovation through AI—the DOE’s Argonne National Laboratory is contributing to projects aimed at strengthening US technological leadership and competitiveness.

One of those projects is the Synergistic Neutron and Photon Science – Intelligence (SYNAPS-I) platform, which integrates data from neutron, X-ray, and microscopy experiments across national labs into a single model. It analyzes information across scales and accelerates understanding of complex systems in real time.

The project is led by Alexander Hexemer, a senior scientist at Lawrence Berkeley National Laboratory (LBNL). SYNAPS-I is a public–private partnership uniting Argonne with LBNL, Brookhaven National Laboratory, Oak Ridge National Laboratory (ORNL), SLAC National Accelerator Laboratory, university researchers, and AI leaders with industry innovators.

Built to speed breakthroughs in microelectronics, medicine, advanced manufacturing, and energy security, SYNAPS-I advances DOE labs with next-generation, AI-driven research capabilities.

“SYNAPS-I is envisioned not just as a tool for analysis and automation, but as a cognitive partner for scientists—capable of generating hypotheses, detecting subtle correlations, and helping turn DOE facilities into truly intelligent, self-driving laboratories,” said Mathew Cherukara of the Argonne SYNAPS-I team.

The project aims to develop an AI-driven imaging engine that turns vast scientific data streams into rapid insight. It seeks to train a multimodal, billion-parameter foundation model on data from more than 100 beamlines across seven DOE facilities, far beyond today’s archive of 50 billion images. “Multimodal” means the model can process different types of data, such as text and images, and “billion-parameter” refers to the internal variables it adjusts as it learns.

Beamlines are experimental stations that deliver and shape X-ray beams for scientific measurements.

To build and test the platform, the team started with ptychography, an X-ray technique that gathers overlapping diffraction patterns—the ways X-rays scatter after interacting with a material—and reconstructs them into sharp, high-resolution images.

"The use of ptychography is expanding rapidly, driven by major light source advances such as Argonne’s Advanced Photon Source (APS) upgrade and the Advanced Light Source (ALS) upgrade at Berkeley Lab,” said Alec Sandy, Associate Director of Argonne’s X-ray Science Division. “Converting raw ptychography data into human and AI-interpretable results in real time maximizes DOE’s investment and makes the measurements immediately relevant for technology development.”

Researchers chose ptychography because it “feels almost magical,” said beamline scientist and SYNAPS-I team member Tao Zhou at Argonne’s Center for Nanoscale Materials (CNM). “Scientists have pushed traditional X-ray optics to their physical limits. Ptychography sidesteps those limits by using physics and computational reconstruction to achieve detail finer than the beam itself can reveal.”

That level of resolution has been achieved before, but not this quickly. SYNAPS-I accelerates the entire workflow, delivering high-resolution images fast enough to keep pace with experiments and surpass the limits of conventional optics.

The platform uses computing resources from the Argonne Leadership Computing Facility (ALCF) and the National Energy Research Scientific Computing Center at LBNL. At Argonne’s APS—the world’s brightest synchrotron X-ray source—a coherent beam scans nanoscale samples such as microelectronics and other manufacturing-relevant materials. SYNAPS-I captures the resulting diffraction patterns and reconstructs them into high-resolution images in real time.

“SYNAPS-I is a rapid-analysis method that delivers insights at the pace data is generated, compressing hours or days of analysis into seconds,” said Aileen Luo, lead developer of the SYNAPS-I model for ptychography.

Behind that speed is an AI platform designed to mirror the physics of the imaging tools.

“By building the physics of coherent imaging directly into the model, we’re giving AI the same knowledge a scientist would use,” said SYNAPS-I team member Emon Dey. “That built-in understanding makes it far more accurate and efficient when handling the massive data volumes produced at DOE facilities.”

The platform works across domains, speeding progress in microelectronics, biomedical research, advanced manufacturing, and energy security. It cuts imaging analysis from years to days and enables real-time, AI-driven materials design for next-generation US manufacturing, with potential economic gains from reduced delays, fewer bottlenecks, and faster innovation.

Argonne recently tested the method by running the full SYNAPS-I workflow on microelectronics and quantum samples at a shared APS/CNM beamline. The platform captured data and displayed imaging results instantly at the beamline, while saved data was sent to the ALCF for further refinement.

“The test opened the door to realtime identification of defects in materials, for example, to guide manufacturing processes and enable autonomous discovery campaigns to discover new technologically impactful materials,” said Sandy. Autonomous discovery campaigns are largely self-driving research efforts in which AI systems help design experiments, analyze results, and determine next steps.

Results from the 26-ID beamline showed ptychography performance 10× higher in resolution and contrast and 100× faster than similar experiments without AI workflows. SYNAPS-I analyzed 1.3 terabytes of data on a single GPU in real time; without AI, a comparable experiment would take about 2500 GPU hours.

As the APS expands its coherent imaging capabilities, the team plans to deploy SYNAPS-I more broadly across DOE light-source and neutron facilities. The capabilities under development could support 10 APS beamlines and many more across the DOE complex.

The team also plans to extend SYNAPS-I beyond ptychography, add partners, test it in real experimental settings, and refine it as it scales.

Argonne National Laboratory has published a video on the development and potential of SYNAPS-I.