And machine learning is the solution
Physicists, biologists and chemists use X-ray pulses created by free-electron lasers, known as XFELs, to probe the structures and interactions of molecules. However, XFELs are inherently unstable because of fluctuations in properties of the accelerators they are driven through. This means there are variations between the X-rays created by different pulses, which can in turn lead to inaccuracies in the measurements of the material scientists are studying.
Although researchers can measure the actual properties of the X-rays, many of the existing methods can’t be run at the same time as the experiment because they will interfere with the kit or aren’t fast enough to be run concurrently.
According to a paper published in Nature Communications on 5 June, the problem will only be exacerbated by the next generation of XFELs, such as those under development in Hamburg and the US, which have a much higher pulse rate.
In a bid to tackle the issue, physicists have used machine learning to more accurately predict the properties of X-rays and so make more of the data generated useful to researchers. Their program was trained using datasets from the SLAC National Accelerator Laboratory at Stanford University in the US.
“For current instruments, which generate about a hundred pulses per second, the slow nature of X-ray characterisation means that sometimes up to a half of the data is unusable,” said Alvaro Sanchez-Gonzalez, lead author of the study, which was carried out across 18 research institutions across the world.
“This problem will only be compounded in next-generation instruments… designed to generate hundreds of thousands of pulses per second. Our method effectively resolves the problem, and should work on the new instruments as well as the older ones we tested it on. This will allow useful data to be gathered up to a thousand times faster.”
The researchers say this will allow more detailed analyses of chemical reactions as molecular changes can be observed on timescales of as little as single femtoseconds. The aim is for the method to be installed directly into XFEL instruments so researchers don’t have to apply the program separately. ®