Researchers at the U.S. Department of Energy\u2019s (DOE) Argonne National Laboratory, Lemont, Ill., developed a new technique that accelerates the solving of material structures from patterns uncovered in x-ray experiments. The technique allows researchers to study certain properties, such as corrosion or battery charging and discharging, in real time.<\/p>\n
The technique, called AutoPhaseNN, is based on a method called machine learning, which trains an algorithm on certain experimental data and then uses it to choose the most likely outcome of the current experiment. The data used in this case are created by shining ultrabright x-ray beams from Argonne\u2019s Advanced Photon Source (APS) on a material and capturing the light as they bounce off, a process called diffraction.<\/p>\n
New techniques are important as the APS is in the midst of a massive upgrade, which will increase the brightness of its x-ray beams by up to 500 times. This means that more data will be gathered more quickly once the upgraded APS comes online in 2024, and scientists will need a way to keep up with analysis of that data. Machine learning solutions such as AutoPhaseNN will be a vital part of the more rapid data analyses needed in the future at APS, as well as similar facilities around the globe.<\/p>\n
AutoPhaseNN is an example of an \u201cunsupervised\u201d machine learning, which means that the computer algorithm learns from its own experience how to do a computation more accurately and efficiently, without having to be trained with labeled solutions that have already been figured out, a process which usually involves human intervention.<\/p>\n
\u201cThis new algorithm is essentially able to solve what we call an inverse problem, going from the pieces of the puzzle to create the puzzle itself,\u201d said Argonne computational scientist and group leader Mathew Cherukara, an author of the study published in NPJ Computational Materials. \u201cIn essence, we\u2019re taking a set of observations and trying to identify the conditions that created them. Instead of solving the puzzle by iterating the process of trial and revision based on the prior knowledge, our algorithm assembles the puzzle from the broken pieces in a single step.\u201d<\/p>\n
Getting information about the structure of a material requires scientists to obtain information pertaining not only to the amplitude of the diffracted signal, but also its phase. However, the amplitude, or intensity, is the only part that can be directly measured.<\/p>\n
Because the x-ray beams used to illuminate the sample are coherent \u2014 meaning they all share the same phase initially \u2014 whatever change to the phase occurs as a result of the diffraction can be mapped back to the sample itself, said Argonne nanoscientist and co-author Henry Chan.<\/p>\n
\u201cPhase retrieval is essential for understanding the structure \u2014 most of the relevant information is found in the phase,\u201d said lead author Yudong Yao, an Argonne x-ray physicist at the time of this research. \u201cWith the kind of diffraction we\u2019re doing, getting the phase information is a challenge; it\u2019s like figuring out how all the pieces fit together solely based on the colors you can see on each piece.\u201d<\/p>\n
For conventional, supervised neural networks to solve this inverse problem, the researchers would have had to pair \u201cbroken puzzles\u201d with fully assembled examples so that the neural network could have something to train against. With an unsupervised neural network, the algorithm can learn to stitch together the puzzle from just the broken pieces. The resulting network is fast, accurate and (unlike conventional methods) capable of providing 3D images in real time to scientific users of facilities like the APS.<\/p>\n
<\/p>\n
For more information:<\/p>\n
Argonne National Laboratory<\/p>\n