Seeing the Unseeable: How AI Reveals Atomic-Level Dynamics in Nanoparticles
Observing atomic-level dynamics has been an elusive goal in materials science, limited by physics that forces scientists to choose between seeing fine details or capturing dynamic changes. Now, a breakthrough AI technique developed by CDS Associate Professor of Mathematics and Data Science Carlos Fernandez-Granda and his collaborators has shattered this barrier, allowing unprecedented observations of atoms in motion.
In a paper published in Science titled “Atomic Resolution Observations of Nanoparticle Surface Dynamics and Instabilities Enabled by Artificial Intelligence,” researchers from NYU and Arizona State University have successfully observed atomic-level movements on the surfaces of platinum nanoparticles with millisecond time resolution — a feat previously impossible with conventional microscopy techniques.
“When you try to observe nanoparticles under realistic conditions, you face a fundamental challenge,” said Fernandez-Granda. “If you shoot too many electrons at the material to get a clear image, you destroy what you’re trying to study. But with too few electrons, your images are dominated by noise.”
The breakthrough came through a novel unsupervised deep learning methodology developed by Fernandez-Granda, Peter Crozier, and collaborators to train deep neural networks and also evaluate them using only noisy data. That approach, called “unsupervised deep video denoiser,” has the ability to clean up extremely noisy electron microscope data without requiring clean reference images. For the electron microscopy data, this approach improved the signal-to-noise ratio of the raw data by a factor of 40, revealing atomic-level details previously hidden in the noise.
This research began when Fernandez-Granda met Crozier from Arizona State University at an interdisciplinary National Science Foundation “Ideas Lab” workshop in 2019. Crozier showed him extremely noisy electron microscope images of catalytic nanoparticles that are crucial for environmental applications.
“I thought we’d just apply our existing neural network and it would work,” Fernandez-Granda said. “It didn’t work at all. This started a multi-year project where we had to find ways to make neural networks denoise this data effectively.”
The technical challenges were formidable. Traditional denoising methods require clean reference images for training, but no such images existed for these dynamic nanoparticles. The team created a specialized neural network architecture that could learn from the noisy data itself, building upon recent advances in unsupervised denoising.
What they discovered with this new technique was remarkable. The nanoparticle surfaces constantly shifted between ordered crystalline structures and disordered “adlayers” of rapidly moving atoms. These transitions happened in less than 100 milliseconds at room temperature. Even more surprisingly, these surface instabilities sometimes penetrated deeper into the particles, causing defects like stacking faults.
This work has important implications for catalysis, where materials like platinum nanoparticles are essential for chemical reactions in industrial processes and environmental applications. Understanding their dynamic behavior could lead to better catalyst designs with higher efficiency and longevity.
The research represents an outstanding example of interdisciplinary collaboration. CDS PhD alumnus Sreyas Mohan played a leading role in developing the denoising methodology, while other team members included Courant Instructor & Assistant Professor Matan Leibovich, and former visiting CDS students Adria Marcos Morales and Shreyas Kulkarni.
Beyond catalysts, this approach opens new avenues for studying all kinds of materials at the atomic level under realistic conditions. The combination of advanced electron microscopy with AI-powered image processing promises to reveal dynamic processes previously hidden from scientists. Fernandez-Granda, together with Crozier, Mohan and CDS PhD alumnus Kangning Liu have recently released a tutorial article (with accompanying videos and code) on AI-powered denoising, which provides a self-contained introduction to this methodology.
“Our experience with electron microscopy shows that when you try to make things work in applications that matter, interesting methodological questions arise,” Fernandez-Granda noted. “The machine learning advances we developed to make this work have applications well beyond this specific project.”
The paper is the culmination of multiple publications and years of collaboration between data scientists and materials scientists, demonstrating how interdisciplinary approaches can lead to fundamental scientific discoveries.
By Stephen Thomas
