AI helps microscopes find the most informative nanoscale features

Researchers at the Department of Energy's (DOE) Oak Ridge National Laboratory (ORNL) have developed an artificial intelligence framework that helps researchers use atomic force microscopes to identify important nanoscale features while autonomously targeting the most informative areas of a sample for closer study.

Although atomic force microscopy (AFM) reveals structures as small as molecules, operating the instrument still requires expert judgment about where to scan, how to adjust settings and which features deserve closer study. SimuScan reduces that burden, making AFM faster, more consistent and better suited for high-throughput research.

"Operating an atomic force microscope is a bit like piloting a modern jet," said Liam Collins, an ORNL senior R&D scientist at the Center for Nanophase Materials Sciences (CNMS). "The hardware has incredible capability, but making full use of it often requires an experienced pilot." That reliance on specialized expertise slows large-scale studies and makes results more dependent on individual users.

"The challenge is not only acquiring the image but also understanding what is in the image, deciding what matters and knowing where the microscope should look next," said Ruben Millan Solsona, an ORNL technical professional and staff scientist.

In a paper published in Nature Communications, the researchers describe how SimuScan addresses one of AI's biggest obstacles for AFM: the shortage of high-quality labeled training data.

Why interpreting AFM data is difficult for AI

At first glance, interpreting AFM images looks like a standard image-analysis problem. But AFM images differ fundamentally from photographs because they reflect both the sample and the measurement process.

"A camera records reflected light, but an atomic force microscope acts more like a high-tech record player needle feeling its way across a landscape," Collins said. What the microscope records depends on the sample, the AFM probe itself and the way it is measured.

Millan Solsona put it simply: "Tip geometry, drift, flattening and contamination can all introduce artifacts that resemble real nanoscale structures. Experienced users learn to distinguish them; AI models must be taught to do the same."

Training data are also scarce. Unlike everyday photographs or medical images, relatively few AFM images have been carefully labeled by experts, making it difficult to train AI models.

A flight simulator for AFM

SimuScan tackles the data problem by generating synthetic AFM images, along with automatic labels tied directly to the simulated object geometry, so models can be trained without large volumes of hand-annotated experimental data.

To work in real laboratories, however, the synthetic images must be realistic. Rather than producing pristine images, SimuScan recreates the imperfections AFM users encounter every day - including tip effects, scanner drift, electronic noise, contamination and surface roughness.

The researchers validated SimuScan by training AI models on synthetic images and testing whether they could accurately identify features in real AFM data.

"The true test of realism is whether an AI trained in a flight simulator can successfully land a real plane in a storm," Collins said.

Cutting the data labeling bottleneck

Creating training data is often the most labor-intensive part of scientific AI. Experts may spend days or weeks outlining features, and results can vary from one researcher to another.

With SimuScan, much of that burden shifts to computation. The system can generate large datasets, including thousands of labeled images, with controlled variability in object shapes, backgrounds and artifacts. In this approach, experimental data are used primarily to test and refine the models rather than create most of the training labels.

The guiding insight, Millan Solsona said, is that synthetic data must not be too perfect. Training on realistic imperfections helps AI recognize true nanoscale structures under real laboratory conditions.

Finding targets and deciding where to zoom in

Beyond analyzing images after the scan, SimuScan supports a closed-loop approach to targeted imaging. The process starts with a fast, low-resolution survey scan across a relatively large area. The AI then identifies and segments features, ranks potential targets based on user-defined criteria and directs the microscope to regions most likely to contain scientifically relevant information. The process can repeat as needed.

The autonomy is deliberate but bounded. The scientist still sets the goals - what to look for, what constraints to respect and when enough data have been collected. For researchers using CNMS, this capability could make advanced AFM measurements more accessible, even for scientists who are experts in their materials but not in microscopy.

Collins likens it to moving from a highly skilled pilot to an intelligent copilot. The point is not to remove the researcher from the loop but to shift the person to higher-value decisions: interpreting results instead of hunting for them.

Testing SimuScan in the real world

In demonstrations spanning fabricated nanostructures, DNA assemblies, and bacterial cells, the team showed that models trained primarily on synthetic data could transfer to real AFM images.

The largest sources of error, the researchers found, often came from the background rather than the target. Real materials can have regular nanoscale textures such as atomic terraces, grain structures and periodic patterns that confuse models. Dense clusters or overlapping objects, such as touching bacterial cells, also remain challenging, even for humans.

Those results are helping the team improve SimuScan by better representing realistic substrates so models learn what to ignore as well as what to detect.

Toward routine, scalable AFM

For SimuScan to become commonplace, Collins and Millan Solsona said it must integrate more seamlessly with microscope software and include straightforward ways to verify that AI models remain reliable as instruments and experimental conditions change.

The first applications are likely to involve studies requiring measurements of thousands of similar objects, such as nanoparticles, DNA nanostructures and bacterial cells.

In the long run, Collins and Millan Solsona envision microscopes becoming active partners in discovery rather than passive imaging tools. SimuScan is one step toward that future.

The work was supported by DOE's Office of Science through the Biopreparedness Research Virtual Environment initiative and relied on a user project at CNMS, a DOE Office of Science user facility at ORNL.

UT-Battelle manages ORNL for DOE's Office of Science, the single largest supporter of basic research in the physical sciences in the United States. The Office of Science is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science. - Scott Gibson