Revealing atomic-scale catalyst dynamics without the need for high-performance computing.

The School for Engineering of Matter, Transport and Energy is part of the Ira A. Fulton Schools of Engineering at Arizona State University. With programs in mechanical engineering, chemical engineering, materials science and engineering and aerospace engineering, the school used imaging and electron microscopy to explore the structure, composition, and properties of materials at micro- to atomic scales.

The Challenge

Understanding how catalysts work requires researchers to observe individual atoms as they move, rearrange and grow under realistic operating conditions. Modern direct electron detectors make it possible to capture these processes at millisecond timescales, but doing so introduces a fundamental challenge.

To achieve frame rates of around 10–100 ms per image, only a very small number of electrons can be collected in each frame. The resulting low-dose images are dominated by shot noise, obscuring the subtle atomic-scale features needed to understand catalytic behaviour.

Increasing the electron dose improves image quality but risks damaging the specimen or altering the very processes under investigation. This dose–time trade-off is one of the fundamental limitations of dynamic transmission electron microscopy.

Existing approaches to overcome this limitation include AI models trained using extensive simulated or experimental datasets but they often require significant GPU or high-performance computing (HPC) resources, making them computationally demanding and are not always practical for routine day-to-day experimental workflows.

The Solution

Arizona State University deployed SenseAI Enhance, a post-processing software suite which, rather than relying on extensive training and fine tuning of deep neural networks or microscope-specific models, exploits the spatial and temporal redundancy naturally present within each recorded image sequence to suppress noise and recover fine structural detail. It can work on both static images and video, processing complete datasets locally on a standard workstation. The end result is a rapid reconstruction shortly after acquisition without requiring dedicated AI infrastructure or model training.

Because the approach operates directly on the acquired experimental data in a close-formed solution, it avoids the dependency on extensive parameter tuning and reduces concerns associated with learned image priors.