2/4/2025 | 3:30 PM - 5:30 PM | Regency B

Machine learning approaches for high-resolution transmission electron microscopy data interpretation

Author(s)

Mary Scott | University of California, Berkeley

Abstract

Tranmission electron microscopy (TEM) is an essential tool in materials design, enabling the observation of atomic-scale and microstructural features that significantly influence material performance. Historically, TEM has had a transformative impact on many fields, particularly materials science and biological sciences. By imaging individual atomic coordinates in solids, EM can directly relate structure to function with incredible detail. Current TEM capabilities offer various imaging techniques to investigate a material’s local structure and composition in a vareity of modalities, such as direct real-space imaging, diffraction, and spectroscopy. With resolution reaching sub-Angstrom levels, a single image from a high-resolution electron microscope can measure atomic positions, defects,  strain, and composition. Recent advances in detector speed, automated imaging, and robotic sample preparation are making TEM imaging increasingly high throughput. Integrating these TEM methods for structural validation into high-throughput materials prediction, design, and synthesis workflows could greatly accelerate the materials discovery process. However, TEM image analysis and interpretation continue to pose challenges that can be addressed using machine learning and computer vision approaches. 

Recent developments in machine learning tools for computer vision, particularly convolutional neural networks (CNNs), show great promise for high-throughput analysis of EM data. However, CNNs are typically designed for natural images, so optimizing their performance and minimizing bias when applied to scientific imaging data necessitates careful attention. Factors such as data curation, CNN architecture, and the choice of performance metrics can significantly impact the accuracy of CNN predictions on TEM data. Customizing CNNs for scientific applications can be particularly challenging, especially when dealing with smaller datasets.

Here, we systematically investigate the application of convolutional neural networks (CNNs) to electron microscopy (EM) data. We examine how architectural choices affect image classification and segmentation, highlighting the impact of data curation on the balance between flexibility and accuracy in CNN performance. We then apply these methods to a large-scale synthesis study of catalytic nanoparticles. By varying experimental parameters during synthesis, we create cubic-shaped cobalt oxide nanoparticles with different sizes, degrees of corner truncation, and face convexity. We demonstrate effective strategies for applying CNNs to the hundreds of thousands of nanoparticles imaged in this study and discuss the associated error metrics and interpretation of CNN results. The resulting statistical distributions of the nanoparticles provide insights into how synthetic parameters influence nanoparticle structure and shape, ultimately affecting their catalytic behavior. These findings illustrate how design choices in neural network workflows impact TEM image analysis and offer guidance for researchers using CNNs in scientific imaging. Overall, this work exemplifies the large-scale EM analysis of numerous nanoparticles synthesized under various conditions, enabled by machine learning, marking a significant step towards integrating high-throughput TEM analysis into automated and autonomous nanomaterial synthesis workflows.