Science1 publisher2 min readPublished
ORNL's SimuScan trains microscope AI on deliberately imperfect synthetic scans
Labeled atomic force microscopy images are scarce, so an Oak Ridge team simulated the drift, noise and tip artifacts that operators fight every day, then tested whether models trained on the simulation find real features in real scans.
The Scientist · Science desk
What happened
- An Oak Ridge National Laboratory team described SimuScan in Nature Communications, a framework that generates synthetic atomic force microscope images whose labels come directly from the simulated object geometry.
- The simulator reproduces tip effects, scanner drift, electronic noise, contamination and surface roughness, the everyday flaws that can mimic real nanoscale structure in a scan.
- For targeted imaging, a fast low-resolution survey scan comes first, after which the model segments features, ranks candidate targets against user-defined criteria and directs the microscope to them.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint Moving annotation into computation changes what limits a model: it can only learn the artifacts its generator knows how to make, and a tip behaviour or contamination the simulator omits is a gap that more synthetic images will not close.
- exposure A shared simulator standardizes error as well as skill. Where results used to vary with whoever was at the controls, every instrument running the same artifact model inherits the same blind spot.
- capability Target selection is now a written rule, so another group can rerun the same survey-and-zoom pass, or check afterwards what the criteria were.
- decision Groups adopting this have to re-plan scarce microscope time, since experimental images are now the test set and the hours go to checking the model against their own probes and samples.
An atomic force microscope does not take a picture. It feels a surface with a sharp probe, and what comes back depends on the sample, the probe and the way the measurement was made [8]. "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," said Liam Collins, an ORNL senior R&D scientist at the Center for Nanophase Materials Sciences [9][5].
The confusions are specific. Ruben Millan Solsona, an ORNL technical professional and staff scientist, said: "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" [7][10].
Two shortages sit behind this work. First, labeled data: relatively few AFM images have been carefully labeled by experts [11], and an expert may spend days or weeks outlining features, with results that vary from one researcher to another [16]. Second, judgment during the session, about where to scan, how to adjust the settings and which features deserve a closer look [3]. On that second point Collins said, "The hardware has incredible capability, but making full use of it often requires an experienced pilot" [4]. ORNL says the reliance on specialized expertise slows large-scale studies and makes results more dependent on individual users [6].
SimuScan addresses the labeling shortage head on, by taking each label from the geometry of the simulated object [12]. The system can generate thousands of labeled images with controlled variability in object shapes, backgrounds and artifacts [17]. The expertise question it addresses through ranking, and the ranking criteria come from the user [20].
Everything then rests on how convincing the fake images are. Millan Solsona said the guiding insight is that synthetic data must not be too perfect, because training on realistic imperfections helps a model recognize true nanoscale structures under real laboratory conditions [19]. Collins set the bar for the check: "The true test of realism is whether an AI trained in a flight simulator can successfully land a real plane in a storm" [15].
The test as described asks one question, which is whether models trained on synthetic images can accurately identify features in real AFM data [14]. ORNL's summary goes further, saying SimuScan makes AFM faster, more consistent and better suited for high-throughput research [21]. Feature identification and throughput are separate measurements. The phys.org account describes the first being done, but it does not give an accuracy figure or say how many real images were used [22][23].
What to watch
- The Nature Communications paper's numbers: segmentation accuracy on real scans, how many real images were held out, and which sample types they covered.
- Whether the closed-loop targeting has run unattended on a user-facility instrument, and how its ranked targets compare with an experienced operator's picks.
- Whether the simulator transfers to other probe geometries and imaging modes without being retuned for each instrument.