X-rays may be employed as an ultrafast, atomic-resolution camera, and researchers can get atomic-resolution photographs of a system at two points in time by firing a pair of X-ray pulses only minutes apart. The comparison of these photos reveals how a material swings within a fraction of a second, which might aid scientists in designing future generations of super-fast computers, communications, and other technology.
To resolve the information in these X-ray pictures, Joshua Turner, a principal scientist at the Department of Energy’s SLAC National Accelerator Center and Stanford University, and eleven other researchers used artificial intelligence to automate the procedure. Their machine learning-aided approach reported on October 17 in Structural Dynamics, accelerates and expands this X-ray probing technique to previously inaccessible materials.
“What excites me the most is that we can now access a new spectrum of metrics that we couldn’t previously,” Turner said.
Taking care of the blob
When employing this two-pulse approach to analyse materials, the X-rays reflect off the substance and are normally detected one photon at a time. These scattered photons are measured by a detector and utilised to generate a speckle pattern—a blotchy picture that shows the specific arrangement of the sample at one point in time. To quantify variations in the sample, researchers compare the speckle patterns from each pair of pulses.
“However, every photon causes an electrical charge explosion on the detector,” Turner said. “When there are too many photons, these charge clouds combine to form an unidentifiable glob.” Because of the cloud of noise, the researchers will need to gather massive amounts of scattering data in order to have a good grasp of the speckle pattern.
“A lot of data is required to figure out what’s going on in the system,” said Sathya Chitturi, a PhD student at Stanford University who conducted this research. Turner and coauthor Mike Dunne, director of SLAC’s Linac Coherent Light Source (LCLS) X-ray laser, counsel him. To comprehend the speckle patterns, old approaches required collecting all of the data first, then analysing it using models that predict how photons pack together at the detector—a time-consuming procedure.
The machine learning technique, on the other hand, directly extracts fluctuation information from the raw detector picture of scattered photons. This new technology is ten times quicker on its own and 100 times faster when paired with upgraded hardware, allowing for real-time data processing.
The work of coauthor Nicolas Burdet, an associate staff scientist at SLAC who constructed a simulator that provided data for training the machine learning model, contributed to the novel method’s success. The programme was able to learn how the charge clouds merge and how many photons hit the detector per blob and per pulse pair as a result of this training. Even in blobby circumstances, the model proved to be correct.
Seeing through the clouds
The model can extract data from a variety of materials that have previously been difficult to investigate because X-rays reflect off them too weakly to detect, such as high-temperature superconductors or quantum spin liquids. Chitturi believes the new approach might be extended to non-quantum materials such as colloids, alloys, and glasses.
Turner believes the findings will be useful for the LCLS-II upgrade, which would enable researchers to capture up to a million photos, or a few gigabytes of data, each second, compared to approximately a hundred for LCLS.
“At SLAC, we’re delighted about this update but also concerned about how we’ll manage this much data,” Turner said. The scientists discovered in a companion publication that their new approach should be quick enough to cope with all of that data. “This new algorithm will be quite beneficial.”
The increased speed provided by artificial intelligence has the potential to change the experimental process itself. Instead of making judgments after collecting and analysing data, researchers will be able to assess data and make modifications while collecting data, thus saving time and money spent on the experiment. It will also enable the researchers to detect surprises and real-time reroute their trials to examine unexpected behaviours.
“By allowing you to make judgments about changes in experimental variables such as temperature, magnetic field, and material composition at various points throughout the experiment, you may explore more of the materials science you’re interested in and optimise the scientific effect,” Chitturi added.
The project is part of a wider cooperation involving SLAC, Northeastern University, and Howard University to promote materials and chemistry research using machine learning.