Particle physics research

ALCF Resources Advanced Materials Characterization Research

July 5, 2022 — Using ALCF computing resources, a team led by University of Michigan scientists is advancing tomographic reconstruction algorithms to improve image quality for characterization research materials.

a) As the tomographic experiment progresses, projections are collected over an angular range. The measured projections are fed into the dynamic CS algorithm for 3D reconstruction. b) As the amount of data increases, the root mean square error (RMSE) decreases. c) 2D slices of the 3D reconstruction at different timestamps. Click to improve. Credit: Robert Hovden and Jonathan Schwartz, University of Michigan.

Using advanced imaging techniques, scientists can peer inside chemical systems to explore their structure and behavior at the microscopic level. Such research can provide information that facilitates the search for promising new materials for important applications, ranging from batteries to semiconductors.

An ongoing project at the US Department of Energy’s (DOE) Argonne National Laboratory seeks nothing less than to demonstrate the theoretical limits of matter visualization, that is, to identify the most small features of objects that can be imaged and in what degree of detail.

Leveraging resources from the Argonne Leadership Computing Facility (ALCF) and the Advanced Photon Source (APS), and led by researchers at the University of Michigan, the project aims to overcome experimental limitations and improve the quality image in the characterization of materials. To this end, the team relies on algorithmic advances that enable increased visual resolution, improving scientists’ ability to identify and explore important chemical information. The ALCF and APS are user facilities of the US DOE Office of Science.

Given the prominent role of materials science in disciplines spanning engineering, chemistry and physics, such advances are expected to impact a wide range of technologies, including clean energy essentials, such as photovoltaics.

Algorithmic image reconstruction

Electron and X-ray tomography are techniques for obtaining three-dimensional structures of samples from 2D projection images from electron and X-ray microscopy. Advances in tomographic image reconstruction are likely to generate images of tiny matter that were not possible before. In a paper recently published in npj Computational Materials, researchers establish a boundary-breaking method for creating high-resolution visualizations of atomic structure. The technique fills visual gaps with inferred information by bombarding targeted matter with electrons, producing differences in energy absorption that can serve as a kind of chemical fingerprint.

Specifically, the article introduces multimodal electron microscopy, a technique for obtaining high signal-to-noise ratio (SNR) of material chemistry at nanoscale and atomic-scale resolutions. SNR measures the fidelity of a signal, with a higher ratio providing scientists with a clearer picture of the system.

“Our multimodal approach combines signals from an array of detectors to produce more detailed images than previously possible,” said Robert Hovden, a professor at the University of Michigan and principal investigator (PI) of the supported project. by the ALCF Data Science program.

“The data composite that is harvested can help us recover, with a very high signal-to-noise ratio, the chemistry of any system we are examining. Learning the ratios of the elements involved used to be quite difficult, but now we can identify the individual atoms in a system and assign each a chemical composition. Not only that, we can get the amount of chemical composition i.e. stoichiometry. Stoichiometry specifically refers to the amounts of reactants and products in a chemical compound reaction.

Indeed, the SNR recovery method has been tested against experimental data and computer simulations run on the ALCF’s Theta supercomputer. Stoichiometry can now be determined by exploiting data collected by scattering electrons on a target material.

Hovden’s team worked closely with Argonne staff to develop and test the computational viability of the project’s governance algorithm.

Simulations represent one of two computationally expensive components that are integral to research, the other being multimodal reconstruction.

“Extremely large simulations were necessary. The images generated by these simulations are deceptively simple, when in fact each pixel is assembled from thousands of quantum mechanical calculations,” Hovden said.

More efficient interpretation of signals

The work grew out of a desire to implement diffusion physics techniques on different length scales. In physics, scattering refers to changes in a particle’s trajectory due to a collision with another particle.

“The X-ray community had done this kind of work, and given that the use of electron beams produces scattering physics roughly analogous to that of X-ray length scales, it seemed likely that some of these approaches could apply to the research I was doing,” Hovden said.

Current tomographic methods for interpreting atomic structure from elastic scattering signals fail to adequately describe the chemistry of a given sample.

Similarly, chemical composition is determined using scattering signals derived from electron energy loss and/or energy dispersive X-ray interactions, but current techniques tend to require doses of electrons so large that they produce noise or even damage the sample on which they are used. examine.

“Multimodal electron microscopy, in comparison, significantly improves SNR for chemical maps, with gains on the order of 300-500% frequently demonstrated,” said Huihuo Zheng, computer scientist at ALCF and co-PI of the project. . “Additionally, the method can reduce electron doses by more than an order of magnitude, resulting in reduced environmental disturbance.”

The metrics obtained from the simulations help predict when and to what extent the approach fails.

Beyond its advance in combining detectors and computational techniques to reconstruct particles, the impact of the work is evident at APS.

“Put simply, achieving the higher resolutions we want to achieve will enable imaging of materials that we currently cannot image,” said co-PI Yi Jiang, an Argonne Beamline data scientist. working at APS and co-author of the article. “We have already begun to adapt similar multimodal image processing techniques for X-ray microscopy.”

In the future, Hovden and his team intend to extend their method to three dimensions.

“It really involves two questions,” Hovden said. “First, there is the question of whether this dramatic improvement in SNR that enables visualization of atomic structure can also enable 3D visualization of complete chemistry. If the answer is yes, then we need to find out at what resolution and for what classes of materials this is true.

Source: Nils Heinonen, ALCF