Merging X-Rays, Machine Learning to Observe Battery Processes
Argonne National Laboratory researchers have developed a technique combining X-ray technology and advanced computing to understand material structures such as those involved in battery corrosion.
Investigators from the U.S. Department of Energy’s (DOE) Argonne National Laboratory (ANL) have combined advanced computing and X-ray technology to help define material structures from fragmented atomic structures in a backward-working process, something like piecing together the pieces of a jigsaw puzzle to complete a picture.
Argonne National Laboratory researcher Swati V. Pol performs an X-ray spectroscopy study of a lithium-ion battery system at the APS. Image used courtesy of Flickr
The researchers published their work entitled, “AutoPhaseNN: unsupervised physics-aware deep learning of 3D nanoscale Bragg coherent diffraction imaging,” in NPJ Computational Materials.
Advanced Photon Source Accesses X-ray Technology
The researchers collected data by shining super-bright, high-energy X-ray photon beams onto a material and capturing the light bouncing off it. Depending on the material or experimental setup, light can either be reflected (as it bounces off) or diffracted (which causes light to bend/spread). The researchers used the ANL’s Advanced Photon Source (APS) to access the X-ray technology. The APS is funded by the DOE’s Office of Science-Basic Energy Sciences.
The photon beams projected onto the experimental material are uniform, or occurring in the same phase. Phase refers to the degree of refraction or reflection in a beam. A beam’s phase can be determined by the type of waves it has. The change to the uniformity of light as it bounces back off the material can be analyzed and mapped back to the sample material.
Video used courtesy of ANL
The APS is undergoing an upgrade, which will see the brightness of its X-ray beams increase by 500-fold. The capabilities of the upgraded APS will become accessible in 2024. With this upgrade, data can be retrieved more quickly, and the AutoPhaseNN is one method that can keep pace with incoming data.
AutoPhaseNN Uses Unsupervised Machine Learning
The AutoPhaseNN technique is based on machine learning (ML), a subfield of computer science and engineering that has been around for decades. ML involves using algorithms to analyze and learn from data (in this case, the X-ray data) to make predictions or decisions.
In particular, AutoPhaseNN involves the use of unsupervised ML, which involves training a machine to carry out a task without human supervision. It is a type of ML that does not rely on labeled data for guidance.
The APS at the Argonne National Laboratory. Image used courtesy of APS ANL
This means the AutoPhaseNN can learn to piece together a puzzle of information independently of any fully constructed reference, which would otherwise be needed by conventional supervised neural networks for training.
The researchers say AutoPhaseNN results in a more rapid and accurate network, and provides 3D images in real-time. This can provide scientists with information quickly, pushing forward research.
The scientists believe the new ML-based technique could be used to study battery discharging and charging, as well as the corrosion of battery components in real-time.