New chip technology enables real-time insights from scientific data

Every second, scientific experiments produce a flood of data — so much that transmitting and analyzing it can slow down even the most advanced research. To help scientists better manage this data deluge, researchers at the U.S. Department of Energy’s (DOE) Argonne National Laboratory have developed a new computer chip that rapidly compresses and processes the huge amounts of data generated by advanced X-ray detectors, like those at the Advanced Photon Source (APS), a DOE Office of Science user facility at Argonne. By compressing data right at the source, like shrinking a movie or song to make it easier to send, this technology makes experiments faster, more efficient and more insightful than ever.

When X-rays or electrons hit a sample, detectors capture the resulting signals — much like a digital camera captures light to produce photos. These signals are converted into electrical pulses and then digitized into numbers that computers can process. But with modern detectors, the amount of data generated is enormous. Every frame, even those with little useful information, is sent out for storage and analysis. This can overwhelm computer systems and slow down research, making it harder for scientists to find what matters most.

“Experiments at the APS will benefit significantly from this technology. To fully use the capabilities of the source, we need technology like this.” — Argonne physicist Antonino Miceli

“Our goal is to bring more computing right where the data is generated,” said physicist Antonino Miceli of Argonne and the University of Chicago. ​“In our earlier work, we showed how advanced mathematical techniques could shrink data while keeping the important parts for analysis. Now, using new chip technology and improvements in microelectronics, we’ve built a chip that puts the math right inside the detector. Using data collected at the APS 8-ID beamline, the detector can compress the data instantly as it’s acquired.”

This means scientists can do key calculations directly on the compressed data, without needing to decompress it first. Consequently, they can analyze results and get feedback much faster, even while the experiment is still running.

Guided by data: Chips that learn from experiments

Building on their work, the team has now implemented a fast, compact matrix-math processor into the detector chip itself. Instead of sending every pixel off the instrument, the chip distills each image into a compact set of numbers that preserves the most important features for scientists. The output is always the same size and streams in real time, making it easier to manage and send.

To make the chip even more useful and flexible, it can be customized for each experiment. Before or during an experiment, scientists can upload preset ​“weights” — settings that tell the chip what features to keep. This process is similar to training an artificial intelligence (AI) model. Using sample data, the chip can be programmed to focus on what is most relevant for each experiment.

“In essence, the chips can be trained on what’s most important for the experiment, so it can compress and reduce data on the fly,” explained Tao Zhou, an Argonne scientist who works on the beamline shared by the APS and the Center for Nanoscale Materials (CNM). ​“The hardware is flexible and can be adapted for different types of compression or data reduction such as radial integration.” CNM is a DOE Office of Science user facility at Argonne.

Tests and design studies show this on-chip approach can reduce data by about 100 to 200 times, while running at speeds of up to a million frames per second. That means less data to move, lower power use and fewer cables, making experiments cheaper, more efficient and easier to scale up.

By combining smart data compression with fast hardware, scientists can get answers in real time and adjust their experiments right away. This helps speed up the cycle of discovery and makes the most of every minute at the beamline. The Argonne team is now working to move this chip from the design stage to large-scale fabrication and use in real experiments.

“Experiments at the APS will benefit significantly from this technology,” Miceli said. ​“Often, the detector, not the X-ray source, is the limiting factor. To fully use the capabilities of the source, we need technology like this. This work also shows how collaborations between detector developers and domain scientists can be very impactful.”

The results of this research were published in the Journal of Instrumentation.

Other contributors to this work include Rami Rasheedi, Mohamed Adel Gharib and Salma Abdelzaher (Argonne, University of Illinois Chicago); Nicholas Contini (Argonne, Ohio State University); Mike Hammer and Henry Shi (Argonne, University of Chicago); Senthil Gnanasekaran, Sebastian Strempfer, Tejas Guruswamy, Kazutomo Yoshii and Angelo Dragone (Argonne); Yu-Sheng Chen (University of Chicago); Lorenzo Rota, Dionisio Doering and Angelo Dragone (DOE’s SLAC National Accelerator Laboratory).

This study was funded by the DOE Office of Science, Advanced Scientific Computing Research and Basic Energy Sciences (BES). This work was primarily supported by the AUREIS project, part of Microelectronics Energy Efficiency Research Center for Advanced Technologies, and the Morpheus project, supported by DOE BES/Scientific User Facilities Division’s Accelerator and Detector R&D program.