Researchers at the U.S. Department of Energy’s (DOE) Oak Ridge National Laboratory (ORNL) have developed an artificial intelligence framework designed to help atomic force microscopes (AFM) identify important nanoscale features and autonomously target areas of a sample for closer examination.
The framework, called SimuScan, addresses a key challenge in AFM imaging: operating the instrument and interpreting its data can require significant expert judgment. The system combines synthetic training data with AI-based image analysis and targeted imaging to make AFM measurements more consistent and suitable 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.”
According to Ruben Millan Solsona, an ORNL technical professional and staff scientist, the challenge extends beyond image acquisition.
“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 Millan Solsona.
The researchers detailed the SimuScan framework in a paper published in Nature Communications, focusing on its ability to address the shortage of high-quality labeled data available for training AI models for AFM.
SimuScan Tackles AFM Data Challenges
AFM images differ from conventional photographs because they reflect both the characteristics of a sample and the measurement process. Factors associated with the AFM probe and measurement conditions can affect the resulting image.
“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.
Millan Solsona highlighted the difficulty of distinguishing actual nanoscale features from measurement-related artifacts.
“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.”
The limited availability of expert-labeled AFM images creates an additional challenge for AI development. SimuScan addresses this by generating synthetic AFM images with automatic labels linked to the geometry of simulated objects.
Synthetic AFM Data Supports AI Training
SimuScan is designed to recreate imperfections encountered during AFM measurements rather than generating idealized images. The framework incorporates factors including tip effects, scanner drift, electronic noise, contamination and surface roughness.
The researchers tested the approach by training AI models primarily on synthetic images and then evaluating their ability to 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.
The approach shifts much of the training-data burden from manual expert annotation to computation. SimuScan can generate large datasets containing thousands of labeled images with controlled variations in object shapes, backgrounds and artifacts.
Experimental AFM data can then be used mainly to test and refine the resulting models rather than generate most of the training labels.
AI Targets Areas for Detailed Imaging
SimuScan also extends AI use beyond analysis after an AFM scan. Its closed-loop imaging process begins with a rapid, low-resolution survey of a comparatively large sample area.
The AI identifies and segments features, ranks potential targets according to user-defined criteria and directs the microscope toward regions considered most likely to contain relevant information. The process can be repeated as required.
The system keeps the scientist involved in defining the imaging objectives, constraints and point at which sufficient data have been collected. The approach is intended to allow researchers to focus more on interpreting results while the system handles the process of locating potential targets.
SimuScan Tested on Real AFM Samples
The research team demonstrated the framework using fabricated nanostructures, DNA assemblies and bacterial cells. The tests showed that AI models trained primarily using synthetic data could transfer their capabilities to real AFM images.
The researchers found that backgrounds could be a significant source of errors. Regular nanoscale textures, including atomic terraces, grain structures and periodic patterns, can confuse AI models. Dense clusters and overlapping objects, including touching bacterial cells, also remain challenging.
The findings are being used to improve SimuScan's representation of realistic substrates so that AI models can better distinguish relevant features from background structures.
Toward Scalable Autonomous AFM
For wider adoption, Collins and Millan Solsona said SimuScan will need to integrate more seamlessly with microscope software and provide straightforward methods for verifying AI model reliability as instruments and experimental conditions change.
Potential early applications include studies involving thousands of similar objects, including nanoparticles, DNA nanostructures and bacterial cells.
The researchers envision the technology supporting a shift toward microscopes that can participate more actively in the discovery process. SimuScan represents an approach to autonomous, targeted AFM imaging that combines AI-based interpretation with automated selection of areas for closer examination.
Reference: https://www.azom.com/news.aspx?newsID=65780
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 a
