AI-Guided X-Ray Imaging Lets a Beamline Choose Where to Look Next
A Brookhaven demonstration let software inspect X-ray images as they arrived, select regions for higher-resolution scans and reconstruct ptychographic images at 200 frames per second. The integrated workflow could make scarce beam time more productive, but its reported performance remains facility-controlled rather than independently benchmarked.
A U.S. national-laboratory team has demonstrated an X-ray imaging workflow that can inspect measurements as they arrive, decide where a beamline should look more closely and reconstruct detailed images fast enough to influence the experiment. The Brookhaven tests move scientific imaging closer to a closed loop in which measurement and analysis guide one another.
The 30-second summary
- What happened? SYNAPS-I software analyzed live X-ray images at Brookhaven's NSLS-II, selected regions for higher-resolution scans and processed microchip data through an integrated imaging pipeline.
- Why does it matter? Beam time is scarce, while modern instruments produce more data than researchers can inspect during an experiment.
- What is the catch? The results come from facility demonstrations. The team has not published an independent benchmark showing that the workflow generalizes across samples, instruments and laboratories.
Letting the data redirect the experiment
Conventional beamline studies often follow a measurement plan set before the final data are visible. Researchers may discover after a session that an important region needed a closer scan, when the allocated beam time has already ended.
At the National Synchrotron Light Source II, or NSLS-II, the SYNAPS-I collaboration tested a different sequence. Software analyzed images as they were collected, identified scientifically interesting regions in a complex sample and sent the instrument back to those locations for higher-resolution measurements. Brookhaven says the system operated for hours without a person choosing every target.
Key Number: 200 images per second. That was the reported reconstruction rate for the real-time ptychography workflow integrated with NVIDIA Holoscan.
Rebuilding a chip from diffraction patterns
The second demonstration used ptychography, an X-ray technique that records overlapping diffraction patterns rather than an ordinary photograph. Computation combines those patterns into a high-resolution image. The method can reveal buried structures inside materials, but iterative reconstruction can become a bottleneck while an experiment is running.
A foundation model trained at Argonne National Laboratory replaced part of that iterative process. Integrated with Holoscan, the pipeline reconstructed images at the reported 200-frame-per-second rate. A separate NSLS-II model rebuilt a three-dimensional image of a microchip from nearly one-tenth of the data normally used for that geometry.
After reconstruction, the team fine-tuned Meta's Segment Anything Model 3 to label materials, structures and regions of interest in the chip volume. That joins acquisition, reconstruction and interpretation in a single workflow rather than leaving analysis until after all measurements are complete.
Why adaptive imaging matters
Synchrotron instruments serve many research groups, and each group receives limited access. A system that surveys a wide area quickly and reserves detailed scans for informative regions could reduce wasted measurements and lower radiation exposure for sensitive samples. Potential uses include locating hidden defects in chips, tracking battery-electrode degradation and examining nanomaterials.
The broader goal resembles other forms of AI-directed laboratory work. NewTqnia previously covered how AI agents changed molecular designs after physical experiments failed. The NSLS-II demonstration applies a related feedback principle to the instrument itself: the latest measurement changes what gets measured next.
Several components, not yet one universal system
SYNAPS-I spans Brookhaven, Argonne, Lawrence Berkeley, Oak Ridge and SLAC national laboratories, with industry partners and shared computing infrastructure. Earlier demonstrations established real-time ptychographic processing and agent-assisted instrument control. The latest Brookhaven work connects those ideas to autonomous target selection and reduced-data tomography on microelectronics.
The collaboration has released parts of its reconstruction pipeline as open-source software. That improves reproducibility at the code level, but reproducing the full result also requires specialized detectors, beamline controls, trained models and computing hardware.
Before we overstate the result
The report is a laboratory announcement rather than a peer-reviewed comparison of the integrated system. It gives headline rates and data reductions but not a common benchmark against alternative pipelines, error rates for autonomous target selection or performance across diverse specimens. The three-dimensional workflow is not yet fully real time, and the team says broader models and additional imaging techniques remain future work. This is a promising facility deployment, not a general-purpose self-driving laboratory.
What comes next
The next milestone is continuous tomography in which a provisional three-dimensional image improves while new projections arrive. Researchers also need to show that automated choices preserve scientifically important regions rather than favoring features the model already recognizes. Success will depend less on one speed figure than on whether multiple facilities can reproduce the workflow, quantify its errors and let visiting scientists use it reliably.
Verified topics and entities
Sources and citations4 sources
External references used to support the reporting in this article.
Published by
NewTqnia Artificial Intelligence Desk
An institutional editorial team within NewTqnia
