Scientists Just Watched Atoms Move for the First Time Using AI

Scientists have developed a groundbreaking AI-driven technique that reveals the hidden movements of nanoparticles, essential in materials science, pharmaceuticals, and electronics.

By integrating artificial intelligence with electron microscopy, researchers can now visualize atomic-level changes that were previously obscured by noise. This breakthrough enables a clearer understanding of how these tiny particles behave under various conditions, potentially revolutionizing industrial processes and scientific discoveries.

AI and Electron Microscopy Illuminate Nanoparticle Behavior

Scientists have developed a new method to reveal how nanoparticles move and change over time. These tiny particles play a crucial role in industries like pharmaceuticals, electronics, and energy. The breakthrough, published in Science, combines artificial intelligence with electron microscopy to create detailed visuals of how nanoparticles react to different conditions.

"Nanoparticle-based catalytic systems have a tremendous impact on society," explains Carlos Fernandez-Granda, director of NYU's Center for Data Science and a professor of mathematics and data science, one of the paper's authors. "It is estimated that 90 percent of all manufactured products involve catalytic processes somewhere in their production chain. We have developed an artificial-intelligence method that opens a new window for the exploration of atomic-level structural dynamics in materials."

Combining AI and Electron Microscopy for Unprecedented Detail

The research, conducted in collaboration with scientists from Arizona State University, Cornell University, and the University of Iowa, merges electron microscopy with AI. This powerful combination allows scientists to observe molecular structures and movements — down to a billionth of a meter — with unprecedented detail and speed.

"Electron microscopy can capture images at a high spatial resolution, but because of the velocity at which the atomic structure of nanoparticles changes during chemical reactions, we need to gather data at a very high speed to understand their functionality," explains Peter A. Crozier, a professor of materials science and engineering at Arizona State University and one of the paper's authors. "This results in extremely noisy measurements. We have developed an artificial-intelligence method that learns how to remove this noise—automatically—enabling the visualization of key atomic-level dynamics."

Revealing Atomic Movements with Deep Learning

Observing the movement of atoms on a nanoparticle is crucial to understand functionality in industrial applications. The problem is that the atoms are barely visible in the data, so scientists cannot be sure how they are behaving—the equivalent of tracking objects in a video taken at night with an old camera. To address this challenge, the paper's authors trained a deep neural network, AI's computational engine, that is able to "light up" the electron-microscope images, revealing the underlying atoms and their dynamic behavior.

"The nature of changes in the particle is exceptionally diverse, including fluxional periods, manifesting as rapid changes in atomic structure, particle shape, and orientation; understanding these dynamics requires new statistical tools," explains David S. Matteson, a professor and associate chair of Cornell University's Department of Statistics and Data Science, director of the National Institute of Statistical Sciences, and one of the paper's authors. "This study introduces a new statistic that utilizes topological data analysis to both quantify fluxionality and to track the stability of particles as they transition between ordered and disordered states."

Reference: "Visualizing nanoparticle surface dynamics and instabilities enabled by deep denoising" by Peter A. Crozier, Matan Leibovich, Piyush Haluai, Mai Tan, Andrew M. Thomas, Joshua Vincent, Sreyas Mohan, Adria Marcos Morales, Shreyas A. Kulkarni, David S. Matteson, Yifan Wang and Carlos Fernandez-Granda, 27 February 2025, Science.
DOI: 10.1126/science.ads2688

The research was supported by grants from the National Science Foundation (OAC-1940263, OAC-2104105, CBET 1604971, DMR 184084, CHE 2109202, OAC-1940097, OAC-2103936, OAC-1940124, DMS-2114143).