Oak Ridge National Laboratory Deploys AI to Automate Atomic Force Microscopy

Researchers at the U.S. Department of Energy's Oak Ridge National Laboratory (ORNL) have deployed a new artificial intelligence framework that allows atomic force microscopes to autonomously identify and target nanoscale features. The system, named SimuScan, shifts the burden of navigating the microscope from a human operator to a machine-learning model, enabling the instrument to scan millimeter-scale areas and zoom in on regions of interest without manual intervention. The development, announced on September 9, 2026, marks a transition for nanoscale imaging from a labor-intensive manual process to a high-throughput, automated pipeline.[1][2][3]

Atomic force microscopy (AFM) is a foundational tool in materials science and biology, capable of mapping surfaces at resolutions down to fractions of a 1-nanometer scale. Unlike optical microscopes that use light, an AFM drags a physical probe—often just a few atoms wide at its tip—across a sample, measuring forces in the 10 to 100 piconewton range to build a three-dimensional topographical map. However, operating the instrument requires constant expert judgment to adjust scan parameters, avoid damaging the delicate tip, and distinguish genuine structures from experimental artifacts.[2][3]

"Operating an atomic force microscope is a bit like piloting a modern jet," said Liam Collins, a senior research and development scientist at ORNL's Center for Nanophase Materials Sciences. "The hardware has incredible capability, but making full use of it often requires an experienced pilot." This reliance on human expertise has historically bottlenecked large-scale studies, limiting researchers to analyzing small batches of samples rather than the thousands of objects required for statistically robust materials discovery.[1][2]

Automating the process with artificial intelligence has proven difficult because of a severe shortage of training data. Training deep learning models like YOLOv8 or U-Net typically requires thousands of manually annotated images. In the context of AFM, generating this data means an expert must spend weeks of expert time painstakingly labeling individual molecules, bacterial cells, or surface defects in existing scans. Furthermore, real AFM data inherently blends the physical characteristics of the sample with the specific geometry of the probe used to measure it, making it difficult for an AI trained on one dataset to generalize to another.[2][3]

SimuScan bypasses the manual labeling bottleneck entirely by training the AI on synthetic data. The framework generates up to 10,000 physically faithful, simulated AFM images of arbitrary nanoscale objects, complete with their corresponding ground-truth masks. Crucially, the researchers programmed the generator to include common experimental imperfections, rather than producing pristine, idealized topographies.[1][3]

"Tip geometry, drift, flattening and contamination can all introduce artifacts that resemble real nanoscale structures," said Ruben Millan Solsona, an ORNL staff scientist and lead developer of the framework. "Experienced users learn to distinguish them; AI models must be taught to do the same." By embedding simulated thermal noise, line-flattening distortions, and surface debris into the training data, the synthetic images prepare the AI for the messy reality of a physical laboratory.[2][3]

To validate the approach, the ORNL team deployed the SimuScan-trained models on physical microscopes analyzing real biological and material samples across areas as large as 1 square millimeter. Without requiring any human-annotated experimental data for training, the vision system successfully located, segmented, and analyzed DNA assemblies, bacterial cells, and fabricated nanostructures in real time. The AI demonstrated the ability to recursively act on its findings, adapting scan parameters and acquiring new high-resolution images of targeted features autonomously.[1][3]

The immediate applications for SimuScan will focus on high-throughput studies that require the measurement of thousands of similar objects, such as characterizing the uniformity of nanoparticle batches or mapping the distribution of proteins on a cell membrane. By integrating the AI directly into the microscope's control software, the framework establishes a closed-loop system where the instrument actively searches for scientifically relevant information rather than passively recording a predefined grid.[2][3]

The shift toward autonomous operation fundamentally changes how researchers interact with nanoscale instrumentation. As the models refine their ability to handle overlapping objects and increasingly complex backgrounds, the role of the human scientist moves away from piloting the hardware and toward interpreting the aggregated data the system produces.[4]