This article explains what atomic force microscopy is, why it has been difficult to automate with AI, and how SimuScan addresses these challenges through synthetic data generation and intelligent, closed loop scanning.
1. What Is Atomic Force Microscopy and Why Does It Matter?
Atomic force microscopy, commonly known as AFM, is a technique used to image surfaces at extremely small scales, down to the level of individual molecules. Unlike traditional microscopes that rely on light or electrons, an AFM uses a physical probe that moves across a sample's surface, measuring tiny variations in height and force.
This makes AFM valuable across many scientific fields, including materials science, nanotechnology, structural biology, and microbiology. Researchers use it to study nanostructures, DNA assemblies, bacterial cells, and other features that are far too small to observe with conventional imaging tools.
Despite its capabilities, AFM has one significant limitation. Getting high quality, meaningful results depends heavily on the expertise of the person operating the instrument.
2. Why Operating an AFM Requires Expert Skill
Using an AFM effectively is not as simple as pointing it at a sample and pressing a button. The operator has to decide where on the sample to scan, how to adjust instrument settings, and which features are worth studying more closely.
Liam Collins, a senior R&D scientist at ORNL's Center for Nanophase Materials Sciences, compared this process to piloting a modern jet. The hardware offers enormous capability, but making full use of it typically requires an experienced operator who understands both the instrument and the material being studied.
This dependency on specialized expertise creates a practical bottleneck. Large scale studies slow down when only a small number of trained users can operate the equipment efficiently, and results can vary depending on who is running the scan. As Ruben Millan Solsona, an ORNL technical professional and staff scientist, explained, the real challenge is not just capturing an image. It involves understanding what is inside that image, deciding what matters scientifically, and knowing where the microscope should focus next.
This is precisely the problem that SimuScan was designed to solve.
3. The Core Challenge: Why AI Struggles With AFM Data
At first glance, teaching AI to interpret AFM images might seem similar to standard image recognition tasks. In practice, it is far more complicated.
3.1 How AFM Images Differ From Photographs
A traditional camera captures reflected light to form an image. An AFM works differently. Collins described it as being more like a high tech record player needle feeling its way across a landscape rather than capturing a picture in the conventional sense.
What an AFM records depends on multiple factors working together, including the sample itself, the physical probe used to scan it, and the specific way the measurement is taken. This means that two scans of similar samples can look different depending on subtle variations in the process.
- Tip geometry can distort how surface features appear
- Drift during scanning can shift or blur fine details
- Flattening and contamination can introduce artifacts that resemble real structures
Millan Solsona pointed out that these factors can all introduce artifacts that closely resemble genuine nanoscale structures. Experienced researchers learn to recognize these artifacts and tell them apart from real features. For an AI model to be useful, it needs to learn the same distinction.
3.2 The Data Labeling Problem in Scientific AI
Modern AI models typically learn by studying large volumes of labeled examples. In fields like everyday photography or medical imaging, millions of labeled images already exist, which makes training AI models comparatively straightforward.
AFM does not have this advantage. Very few AFM images have been carefully labeled by experts, largely because labeling requires specialized knowledge and significant time investment. This scarcity of high quality training data has been one of the biggest obstacles preventing AI from being reliably applied to AFM research.
SimuScan was built specifically to address this bottleneck.
4. What Is SimuScan and How Does It Work?
SimuScan is an AI framework developed by ORNL researchers that generates synthetic AFM images along with automatic labels tied directly to the geometry of the simulated objects. Instead of relying on scientists to manually annotate thousands of real world images, SimuScan creates realistic training data computationally.
The research behind SimuScan was published in Nature Communications, describing how the framework tackles one of the most persistent obstacles in applying AI to atomic force microscopy.
4.1 Generating Synthetic Training Data
Rather than producing clean, idealized images, SimuScan intentionally recreates the imperfections that AFM users encounter during real experiments. These include tip effects, scanner drift, electronic noise, surface contamination, and general surface roughness.
This design choice reflects a deeper understanding of how AI models generalize. A model trained only on flawless, artificial images may fail when it encounters the noisy, imperfect data typical of real laboratory conditions. By simulating realistic flaws from the start, SimuScan prepares AI models for the unpredictability of actual research environments.
4.2 Why Imperfect Data Matters More Than Perfect Data
Millan Solsona summarized the guiding principle behind SimuScan clearly. Synthetic data must not be too perfect. Training models on realistic imperfections helps them recognize true nanoscale structures under real world laboratory conditions, rather than only performing well on artificially clean datasets.
To validate this approach, the ORNL team trained AI models using synthetic images and then tested whether those models could accurately identify features in real AFM data. Collins summarized the logic behind this testing method by comparing it to aviation training. The true test of realism, he said, is whether an AI trained in a flight simulator can successfully land a real plane in a storm.
5. How SimuScan Cuts Down the Labeling Bottleneck
One of the most resource intensive parts of building scientific AI systems is manual data labeling. Experts often spend days or even weeks outlining features in images by hand, and the results can vary from one researcher to another depending on interpretation.
SimuScan shifts much of this burden away from human labor and toward computation. The system can generate large datasets, including thousands of labeled images, with controlled variability in object shapes, background textures, and artifact types.
In this approach, experimental data still plays an important role, but its primary purpose shifts. Instead of being the main source of training labels, real world AFM images are used mainly to test and refine the models that were initially trained on synthetic data. This reduces dependency on scarce, expert labeled datasets while still ensuring the models remain grounded in real world accuracy.
6. How SimuScan Enables Autonomous, Targeted Imaging
Beyond simply analyzing images after a scan is complete, SimuScan supports a more advanced, closed loop approach to imaging that actively guides the microscope during the research process.
6.1 The Closed Loop Scanning Process
The workflow begins with a fast, low resolution survey scan across a relatively large area of the sample. Once this initial scan is complete, the AI identifies and segments relevant features within the image.
From there, SimuScan ranks potential targets based on criteria defined by the researcher and directs the microscope toward the regions most likely to contain scientifically meaningful information. This process can repeat multiple times, allowing the system to progressively refine its focus as it gathers more data.
This closed loop method mirrors how an experienced human operator might work, starting broad and then zooming in on the most promising areas, but it performs this process more consistently and without requiring constant manual input.
6.2 Human Oversight and Bounded Autonomy
Although SimuScan introduces meaningful autonomy into the imaging process, this autonomy is intentionally limited. The scientist remains responsible for setting the overall goals of the study, including what features to look for, what constraints the system should respect, and when enough data has been collected.
Collins described this relationship as shifting from a highly skilled pilot to an intelligent copilot. The goal is not to remove the researcher from the process, but to shift their focus toward higher value tasks such as interpreting results, rather than spending time manually hunting for relevant features.
For researchers at ORNL's Center for Nanophase Materials Sciences, this shift could make advanced AFM measurements more accessible, particularly for scientists who have deep expertise in their materials but limited experience operating specialized microscopy equipment.
7. Real World Testing: What the Results Show
To evaluate SimuScan's effectiveness, the research team tested it across several practical applications, including fabricated nanostructures, DNA assemblies, and bacterial cells. The results showed that models trained primarily on synthetic data were able to successfully transfer their learning to real AFM images.
7.1 Where Errors Still Occur
Interestingly, the researchers found that the most common source of error came from the background of an image rather than the target feature itself. Real materials often contain regular nanoscale textures, such as atomic terraces, grain structures, and periodic patterns, which can confuse AI models trying to distinguish meaningful features from background noise.
Dense clusters or overlapping objects, such as bacterial cells that are touching one another, also proved challenging for the models. Notably, this type of ambiguity is difficult even for trained human researchers to resolve consistently.
These findings are directly informing ongoing improvements to SimuScan. By better representing realistic substrates in its synthetic training data, the framework aims to help models learn not only what to detect, but also what to ignore.
8. What This Means for Researchers
For scientists working in nanotechnology, structural biology, and materials science, SimuScan represents a meaningful step toward making advanced microscopy more scalable. Instead of requiring every researcher to become a microscopy specialist, tools like SimuScan could allow scientists to focus more on interpreting scientific results and less on the technical mechanics of operating complex instruments.
This shift could be especially valuable for studies that require measuring large numbers of similar objects, such as nanoparticles, DNA nanostructures, or bacterial populations, where consistency and throughput matter as much as precision.
9. Challenges and Future Directions
For SimuScan to become a widely used tool, researchers note that it will need tighter integration with existing microscope software systems. There is also a need for clear, reliable methods to verify that AI models continue to perform accurately as instruments age, conditions change, or new types of samples are introduced.
Looking further ahead, the research team envisions a future where microscopes function as active partners in scientific discovery, rather than passive tools that simply capture images. SimuScan is described as one early step toward that broader vision.
10. Conclusion
SimuScan reflects a broader shift happening across scientific research, where AI is increasingly used not to replace expertise, but to extend it. By generating realistic synthetic training data and enabling closed loop, targeted imaging, the framework helps address two of the biggest obstacles in applying AI to atomic force microscopy: the scarcity of labeled data and the need for expert judgment during scanning.
As tools like SimuScan continue to mature, they point toward a future where microscopes are not just passive instruments, but active collaborators in scientific discovery, helping researchers focus their time and expertise where it matters most.