Your phone quietly denoises your night photos before you even see them, politely pretending the sensor did not just panic in the dark like a raccoon in a flashlight. Kang and colleagues are doing a much more extreme version of that idea: using deep-learning denoising to watch gold nanocrystals wiggle, disorder, heal, and dissolve in liquid at atomic resolution, millisecond by millisecond.

That is not a normal thing to see. Atoms usually get the privacy settings of celebrities entering a restaurant through the back door.

The Tiny Gold Soap Opera

The paper, Visualizing Millisecond Atomic Dynamics of Nanocrystals in Liquid, focuses on gold nanocrystals sitting in reactive liquid environments. Nanocrystals are little chunks of ordered atoms. Think of them as microscopic apartment buildings where every atom has an assigned unit, except the building is floating in chemical soup and the tenants keep rearranging the furniture.

The researchers used liquid-cell electron microscopy, a technique that lets scientists image materials while they are in liquid instead of dried out and frozen into a more microscope-friendly pose. Regular transmission electron microscopy is already a marvel: it sends electrons through very thin samples to reveal details far smaller than light microscopes can see. Liquid-cell EM adds a tiny sealed chamber, usually with thin windows, so the sample can stay wet while the microscope does its electron-beam interrogation routine.

The catch? Fast atomic movies are noisy. If you record slowly, you get cleaner images but miss the action. If you record quickly, you catch the action but the images look like they were faxed through a thunderstorm. This is where deep learning enters, wearing a lab coat over its hoodie.

Deep Learning as the World’s Pickiest Window Cleaner

The team combined millisecond-speed liquid-cell EM with deep-learning denoising. In plain English: they filmed the nanocrystals fast enough to catch quick atomic changes, then used a neural network to clean the footage without smearing away the meaningful details.

This matters because scientific denoising is not the same as making your vacation photo less grainy. In science, “pretty” is a trap. A denoising model that invents atoms is not helpful; it is basically Photoshop with tenure. Recent work like SHINE, a self-supervised denoising framework for electron microscopy, shows why the field is moving toward methods that can learn from noisy microscope data without needing perfect clean examples, because perfect clean examples are often imaginary unicorn paperwork.

Tools like combb2.io use related ideas in everyday image enhancement - upscale, denoise, deblur - but here the stakes are less “make the photo crisp” and more “please do not hallucinate a gold atom into existence.”

What They Saw: Order, Disorder, Repeat

The headline result is that gold nanocrystals did not behave like rigid little golden bricks. Their local crystallinity fluctuated reversibly depending on the surrounding chemical environment. Parts of the crystal could become more disordered, then regain order, like a tiny office where everyone briefly ignores the seating chart and then HR walks in.

These transient atomic arrangements influenced two big behaviors: dissolution kinetics and grain boundary relaxation. Dissolution kinetics is the pace and pathway by which the material breaks down into the liquid. Grain boundaries are the internal borders where differently oriented crystal regions meet, like seams between floor tiles installed by contractors who disagreed about north.

The key point is that fleeting structures were not background noise. They affected how stable and reactive the nanocrystals were. That is a big deal for catalysts, batteries, sensors, drug delivery particles, and other nanomaterials that do their work at surfaces and interfaces. At the nanoscale, the “surface” is not just the outside of the thing. It is the workplace, the kitchen, the loading dock, and occasionally the complaint department.

Why This Is Sneaky Useful

Materials scientists often want to know why a nanoparticle performs well, degrades, reshapes, or fails. But if the relevant atomic structure only exists for milliseconds, older tools can miss it. That is like trying to understand a bar fight by arriving the next morning and interviewing the stools.

This study suggests that some nanomaterial behavior depends on short-lived atomic states that appear only under realistic chemical conditions. If these findings hold across more materials and reactions, researchers could design nanoparticles by paying attention not only to their “before” and “after” structures, but also to their twitchy middle moments.

That could help build more durable catalysts, better corrosion-resistant materials, and nanocrystals whose reactivity is tuned by controlling their liquid environment. But the limitations matter. Electron beams can affect samples, denoising needs careful validation, and gold is only one material system. The next step is not declaring victory from the mountaintop. It is doing the less glamorous but vital work: more controls, more chemistries, more comparisons, and more proof that the AI cleaned the window rather than painting a nicer view on it.

The Bigger AI Angle

This paper is a neat example of AI as a scientific instrument amplifier, not a magic oracle. The neural network is not “discovering chemistry” by itself. It is helping humans extract usable signal from noisy, high-speed measurements. That is the good version of AI in science: less crystal ball, more very caffeinated lab assistant with excellent pattern recognition.

And honestly, that role is underrated. A lot of discovery depends on seeing the thing clearly enough before it vanishes. Here, the thing is a gold nanocrystal briefly losing and regaining atomic order in liquid, which sounds like a spa treatment designed by quantum mechanics.

References

1. Sungsu Kang et al., “Visualizing Millisecond Atomic Dynamics of Nanocrystals in Liquid,” Journal of the American Chemical Society, 2026. DOI: 10.1021/jacs.6c08238. arXiv: 2603.24776. PMID: 42314043.

2. Joodeok Kim et al., “Self-supervised machine learning framework for high-throughput electron microscopy,” Science Advances, 2025. DOI: 10.1126/sciadv.ads5552. PMCID: PMC11963987.

3. “Visualizing nanoparticle surface dynamics and instabilities enabled by deep learning denoising,” Science, 2025. DOI: 10.1126/science.ads2688.

4. Sungsu Kang, Minyoung Lee, Jinho Rhee, and colleagues, “Advances in Transmission Electron Microscopy to Image Nanocrystals,” Annual Review of Physical Chemistry, 2026. DOI: 10.1146/annurev-physchem-082624-082524.

5. Eva M. van Ravenhorst et al., “Toward sub-second solution exchange dynamics in flow reactors for liquid-phase transmission electron microscopy,” Nature Communications, 2024. DOI: 10.1038/s41467-024-46842-3.

6. Background: Liquid-Phase Electron Microscopy, Transmission Electron Microscopy, and Convolutional Neural Network.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.