200 images per second: Smart X-ray tomography maps microchip in 3D, no human needed

X-ray experiments at Brookhaven National Laboratory have begun to steer themselves. Machine learning picks which parts of a sample to examine in detail, rebuilds 3D images as data arrive and, in one test, did the job with nearly one tenth of the usual data.

The demonstrations at the National Synchrotron Light Source II (NSLS-II), a U.S. Department of Energy user facility, are early results from SYNAPS-I, short for Synergistic Neutron and Photon Science – Intelligence. It is a seed project under the ModCon initiative in the DOE’s Genesis Mission, led by Lawrence Berkeley National Laboratory’s Advanced Light Source with Argonne, Oak Ridge, SLAC and Brookhaven. The collaboration combines open-source models, industry partnerships and advanced computing resources.

Letting the data pick the target

At NSLS-II’s Hard X-ray Nanoprobe, an X-ray microscope for 2D and 3D nanoscale imaging, a team led by lead beamline scientist Hanfei Yan ran an automated experiment that analyzed images as they were acquired. It flagged scientifically interesting regions of a complex sample and picked those spots for follow-up high-resolution scans, with no human intervention.

“This enables an automated experiment to run for hours, surveying a large area, and collect high-fidelity data only from scientifically important locations,” said Daniel Allan, who leads NSLS-II’s Data Engineering Group. Conventional methods can’t do this reliably, he said.

A microchip in 3D, in real time

A second test focused on ptychographic tomography, a computational technique that builds detailed 3D images from many individual measurements, using microelectronic chips as samples. A foundation model trained at Argonne replaced the slow iterative process normally used to reconstruct the images. Working with NVIDIA and its Holoscan platform, the team reconstructed images in real time at 200 images per second.

A model developed and trained at NSLS-II reconstructed a microchip from nearly one-tenth of the data normally required, which helps with thin, flat samples where complete tomographic datasets are hard to collect. The team then fine-tuned Meta’s open-weight Segment Anything Model 3 for chip imaging so it could automatically pick out materials, structures and regions of interest in the 3D volume. Together, the steps can cut scan time and radiation dose while speeding up interpretation.

Why it matters beyond chips

Yan said the technique could help any field that needs fast, high-resolution X-ray tomography and rapid 3D data analysis. Examples include finding buried structures and defects in microelectronics, studying degradation in battery electrodes and characterizing complex nanomaterials. The tools also help users make better use of limited beam time, he said.

Tomography while the experiment runs

The team plans to train more robust foundation models on more diverse datasets and extend the approach to other 3D imaging techniques. A key goal is real-time tomography. Today, researchers usually collect an entire projection dataset before reconstructing it. In a real-time workflow, reconstruction would start during the experiment, with a low-resolution image sharpening as more measurements arrive. Paired with automated feature detection, that would let beamlines adapt on their own to what they see.

Sean McSweeney, NSLS-II’s interim deputy for strategy, said such workflows will be integral to next-generation user facilities. Alexander Hexemer, the SYNAPS-I lead contact at the Advanced Light Source, said facilities share the same challenges, from data volumes that outpace human analysis to reconstruction bottlenecks, and that no one facility can solve them alone.

As these technologies mature, the goal is not simply to process experimental data faster. It is to tighten the link between measurement, analysis, and scientific decision-making, so experiments can respond to what they discover as they run.