Deep Learning to Understand Heterogeneous Catalysts
A deep learning approach allows researchers to rapidly analyze three-dimensional representations of multi-component catalysts
The Science
To obtain three-dimensional (3D) representations of materials, researchers use transmission electron microscopy tomography—a specific type of microscopy that takes images from different projections. These projections are reconstructed into a 3D grayscale image that contains rich information about the structure of complex materials. These 3D images need to be segmented, or divided into regions with similar properties, to obtain quantitative data. Researchers used machine learning (ML) to segment images of a catalytic material based on platinum (Pt) nanoparticles supported on alumina. The ML model automatically determines important structural information with human level precision and on a statistically-relevant scale.
The Impact
Image segmentation is an incredibly time intensive process prone to variation between researchers. New work uses ML to streamline and standardize this necessary component of electron microscopy-based tomography. In the test case, the ML program analyzed images of Pt nanoparticles supported on alumina. The program allowed the researchers to obtain data on the Pt size distribution and Pt-alumina interactions via automation. Streamlining the segmentation process saves time and increases the overall efficiency of studying materials via tomography.
Summary
3D tomography is a powerful tool for studying complex materials but is time intensive for researchers. Analyzing the resulting images requires laborious segmentation that includes human-induced variation. Researchers developed a ML model to automate segmentation of the images of catalytic materials. Researchers chose to study a system of Pt nanoparticles deposited on an alumina support because of its complex and poorly understood structure, as well as high relevance to numerous catalytic applications. Trained on a subset of manually segmented data sets of a Pt-alumina catalyst, the model was then run on full 3D tomograms. The model produced statistically-relevant data on the underlying structure of the alumina catalyst, including the location and type of different edges and faces. It also identified the location, size, and shape of Pt nanoparticles on the alumina. Finally, the model classified the different Pt-alumina interactions and showed their distribution across the overall structure. This was accomplished in less time than traditional manual segmentation. This model demonstrates how ML can be integrated into tomography studies to streamline analysis and reduce the amount of tedious work necessary to obtain statistically significant results.
PNNL Contact
Libor Kovarik, Pacific Northwest National Laboratory, libor.kovarik@pnnl.gov
Funding
The experimental work was conducted in EMSL, a Department of Energy (DOE) Office of Science user facility at Pacific Northwest National Laboratory. Libor Kovarik was supported by the DOE, Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences, and Biosciences.
Published: October 28, 2022
A. Genc, L. Kovarik. H. Fraser. 2022. “A Deep Learning Approach for Semantic Segmentation of Unbalanced Data in Electron Tomography of Catalytic Materials,” Scientific Reports, 12, 16267. [DOI: 10.1038/s41598-022-16429-3]