Brands
Latest top stories
Technology

How do you map a microchip in real time? Let an AI steer the X-rays

19 November 2026

Cerium oxide forming in heated cell at 340 degrees C. (Image: Kaitlin Perry/Argonne National Laboratory)

 

 

An artificial intelligence agent at Argonne National Laboratory has successfully located the boundary between two regions of a semiconductor chip by using live X-ray images to decide where to scan next. Researchers assigned the task using plain-language prompts. Following the trial, the US facility announced a major expansion of the shared AI computing service that underpinned the experiment. The upgraded system, named Tara, provides roughly 100 times the graphics processor memory previously available to users.

Operated for the US Department of Energy outside Chicago by UChicago Argonne, LLC, the laboratory runs large-scale facilities open to researchers from universities and industry. Among them is the Advanced Photon Source, a synchrotron that generates intense X-ray beams to probe the internal structure of materials. The autonomous microscope trial, first reported on 5 October, took place at one of its experimental stations, known as beamlines.

 

How AI guides an X-ray microscope

 

At a synchrotron, visiting teams operate under strict time constraints to gather their data. During semiconductor experiments, a significant portion of that allotted window is often spent simply locating a microscopic feature before detailed analysis can begin. The search becomes particularly difficult when the target occupies a tiny fraction of the sample or varies between individual devices. Generating images in real time allows scientists to narrow their focus and direct high-precision measurements exactly where they are required.

Producing those images quickly presents a major technical hurdle. The Argonne team relied on a technique called ptychography, which records how X-rays scatter across a series of overlapping positions on a sample. Each exposure creates a diffraction pattern, and specialist software combines those patterns to reconstruct a high-resolution image. Conventional reconstruction is computationally heavy, so researchers used Argonne’s PtychoFM model to process measurements on the fly, delivering images fast enough to steer the experiment in real time.

To interpret those visuals, the team integrated Meta’s Segment Anything Model 3 to distinguish key features within each frame. The AI agent then analysed the data and directed the microscope as it honed in on the target boundary. According to Argonne, the platform can execute routine adjustments independently while requesting human intervention when it encounters higher levels of risk or uncertainty. However, the laboratory’s published account does not benchmark the system against an experienced human operator, nor does it provide a success rate across varied chip architectures.

 

What Tara adds to Argonne’s AI inference service

 

Built in partnership with Hewlett Packard Enterprise, Tara comprises 672 computing nodes powered by 2,688 Nvidia Grace Hopper chips. Its expanded memory allows Argonne to host substantially larger AI models while serving a wider pool of scientists. The platform now provides access to more than 50 models, spanning computer vision tools and bespoke systems built for scientific discovery.

The Argonne Leadership Computing Facility designates the platform as an inference service because it runs pre-trained models against incoming experimental data. Rather than building local infrastructure, eligible researchers can connect directly to these models from their own beamline software – a capability the microscope team used to query Meta’s segmentation model mid-experiment. Access is open to the facility's user community, Department of Energy scientists, and teams contributing to the department’s Genesis Mission, a national initiative aimed at embedding AI across scientific research.

 

From chip trials to routine synchrotron experiments

 

The Argonne trial forms part of SYNAPS-I, a multi-laboratory collaboration led by Lawrence Berkeley National Laboratory to test AI-guided experiments at US X-ray and neutron facilities. In a parallel effort at Brookhaven National Laboratory, researchers used live scan images to automatically select regions for higher-resolution imaging, demonstrating in a separate test that ptychographic images could be reconstructed at a rate of 200 per second. Together, these advances illustrate how rapid data processing is transforming static measurements into responsive, real-time investigations.

European facilities are pursuing similar breakthroughs. At the European Synchrotron Radiation Facility in Grenoble, scientists used machine learning to monitor the evolving properties of thin films via X-ray data, feeding those insights directly back into instrument controls. Both initiatives highlight how the underlying scientific goal dictates what an algorithm must recognise, whether tracking a film as it forms or pinpointing an interface buried deep inside a microchip.

For visiting scientists, the ultimate measure of success will be whether AI-guided searches consistently free up precious beam time for core discoveries. While Argonne’s demonstration proved the concept on a single semiconductor, testing the system across diverse chip designs will determine how reliably the technology performs, and where human intuition remains indispensable. As computing power such as Tara scales up alongside smarter algorithms, the modern synchrotron is steadily evolving from a passive observation tool into an active, autonomous partner in scientific discovery.