The Tech TL;DR:

• Core Innovation: ORNL’s SimuScan uses a synthetic training data approach to teach AI models how to interpret atomic force microscopy (AFM) data without requiring massive manual image annotation.
• Operational Impact: Mitigates user-dependent operational bottlenecks, reducing the time required to locate and scan high-value nanoscale targets.
• Technical Validation: SimuScan intentionally incorporates real-world scanning imperfections—such as tip geometry, drift, and electronic noise—into synthetic training sets to ensure high-accuracy generalization in physical labs.

Overcoming the Scarcity of Labeled Nanoscale Data

While atomic force microscopy achieves atomic-level structural resolution, operating the hardware demands significant operator expertise to select scan paths, calibrate settings, and identify genuine features versus scanning artifacts. According to Liam Collins, a senior R&D scientist at ORNL’s Center for Nanophase Materials Sciences (CNMS), operating an AFM resembles piloting a modern jet where incredible hardware capability requires an experienced pilot to fully utilize it. Ruben Millan Solsona, an ORNL technical professional and staff scientist, noted that the primary challenge involves not just capturing the image, but determining image contents, deciding significance, and choosing subsequent scanning coordinates.

Training machine learning models to automate this workflow has historically hit a major roadblock: a severe lack of labeled training data. Unlike standard consumer photographs or medical imaging libraries, AFM datasets lack large volumes of expert annotations. Furthermore, AFM data does not reflect simple reflected light like a camera. Instead, the instrument functions similarly to a high-tech record player needle feeling its way across a physical landscape, where measurements depend simultaneously on the sample, the physical probe tip, and environmental factors.

Simulating Imperfections: The Flight Simulator Approach for Microscopes

To bypass the labor-intensive bottleneck of manual data labeling, the ORNL team built SimuScan to generate synthetic AFM images alongside automatic labels directly mapped to simulated object geometries. To ensure these models perform effectively in physical laboratories, the framework deliberately avoids generating pristine, idealized images. Instead, SimuScan simulates everyday operational anomalies including tip effects, scanner drift, electronic noise, contamination, and surface roughness.

As Liam Collins described the validation process, training an AI model on synthetic data and testing it on real-world AFM feeds is comparable to training a pilot in a flight simulator to land a real plane during a severe storm. By baking realistic imperfections into the training pipeline, the AI learns to separate actual nanoscale structures from artifacts caused by tip geometry, drift, flattening, and contamination. Experimental laboratory data is subsequently reserved primarily for testing and refining models rather than generating primary training labels.

Implementation Architecture

Deploying automated synthetic data generation pipelines into physical microscope control software involves establishing programmatic loops that feed simulation outputs directly into neural network training pipelines.

import numpy as np

def generate_synthetic_afm_scan(grid_size=256, drift_factor=0.02):
    # Initialize baseline flat topography
    topography = np.zeros((grid_size, grid_size))
    
    # Simulate nanoscale features (e.g., spherical nanoparticles)
    x, y = np.indices((grid_size, grid_size))
    center_x, center_y = grid_size // 2, grid_size // 2
    topography += np.exp(-((x - center_x)2 + (y - center_y)2) / 200.0)
    
    # Inject realistic scanner drift and electronic noise
    noise = np.random.normal(0, 0.05, size=(grid_size, grid_size))
    drift = np.linspace(0, drift_factor, grid_size)[:, None]
    
    simulated_image = topography + noise + drift
    return simulated_image

# Execute batch generation for training dataset
training_batch = [generate_synthetic_afm_scan() for _ in range(1000)]
print(f"Generated {len(training_batch)} synthetic AFM frames with simulated drift and noise.")

By shifting the burden of data creation from manual annotation to automated computation, frameworks like SimuScan clear the path for scalable, autonomous high-throughput materials characterization.