*Important notice: This news reports on an unedited version of an accepted paper and is awaiting final editing. Therefore, the paper should not be regarded as conclusive or treated as established information.
A machine-learning (ML) model that enhances atomic force microscopy (AFM) image analysis has been introduced recently, bridging the gap between simulated and experimental data. These findings were published in npj Computational Materials.
Study: Improving atomic force microscopy structure discovery via style-translation. Image Credit: Wulan ananda/Shutterstock.com
Using artificial intelligence (AI)-driven unpaired image-to-image translation, the framework transforms simulated AFM images into realistic representations that closely match laboratory measurements. This advancement enables deep learning (DL) models to predict three-dimensional atomic structures directly from experimental images without requiring labeled experimental AFM training data.
Download a copy of this for later!
Challenges in High-Resolution Imaging Techniques
Visualizing atomic structures is crucial for understanding materials at the nanoscale. Frequency-modulated non-contact atomic force microscopy (FM-NC-AFM) with carbon monoxide (CO)-functionalized tips provides sub-nanometer resolution by measuring frequency shifts caused by tip-sample interactions.
However, interpreting these atomic contrast patterns often requires simulations using the probe particle model (PPM) to generate theoretical AFM images. Experimental images are influenced by factors such as instrumental drift, thermal noise, and tip asymmetry.
Consequently, DL models trained solely on simulated images frequently perform poorly on real experimental data due to discrepancies between theoretical and laboratory image characteristics.
CycleGAN: Narrowing the Gap Between Data
To bridge the gap between simulated and experimental AFM images without large labeled datasets, researchers implemented a cycle-consistent generative adversarial network (CycleGAN). This framework employed two ResNet-based generators and two PatchGAN discriminators to learn bidirectional translation between simulated and experimental image domains from unpaired two-dimensional frequency-shift images.
The forward generator transformed simulated AFM images into realistic, experimental-style images that incorporated noise and artifacts. Cycle-consistency and identity losses constrained the translation to reduce unnecessary changes and help preserve structure-relevant image features.
For three-dimensional (3D) structure prediction, the translated AFM image stacks were analyzed by a convolutional neural network (CNN) to estimate atomic positions, followed by a graph neural network (GNN) that reconstructed the molecular graph. The model was trained and evaluated using bilayer water clusters adsorbed on a gold (Au(111)) surface.
Validating the Effectiveness of Style Translation
To assess the quality of style translation, the study trained a CNN-based binary classifier to distinguish simulated AFM images from
experimental ones. This classifier achieved an accuracy of nearly 93.3% on held-out test data. When applied to CycleGAN-generated images, it assigned authenticity scores that closely matched those of real experimental images.
Additionally, metrics such as the Wasserstein distance and the Fréchet Inception distance confirmed that the translated images were significantly closer to the experimental data than images generated with handcrafted perturbations, such as Gaussian noise.
The translated images also enhanced atomic structure prediction. Models trained solely on clean simulated images often failed to identify lower-layer water molecules because experimental noise masked weak height signals.
In contrast, the framework trained on generative style-translated datasets accurately reconstructed atomic positions across both water layers and demonstrated greater robustness to experimental noise.
Since the true atomic structures corresponding to the experimental AFM images are unavailable, the researchers indirectly evaluated the predictions by comparing distributions of local structural properties with theoretical reference distributions derived from simulations.
They evaluated structural properties, including oxygen-oxygen distances, bond angles, hydroxyl orientations, and hydrogen-bond geometry.
Statistical metrics, like energy distance, indicated that models trained on style-translated images produced structures that aligned more closely with DFT predictions than those trained solely on simulated data.
Applications of CycleGAN in Scanning Probe Microscopy
Although the study focused on bilayer water structures adsorbed on an Au(111) surface, the CycleGAN framework can be extended to other scanning probe microscopy applications. By learning common artifacts, it effectively translates idealized simulations into images that resemble laboratory measurements. However, systematic validation is required.
This approach enables scientists to adapt simulated AFM data to various conditions without requiring labeled images, providing a flexible tool for molecular identification and surface structure analysis across diverse materials.
Future Directions for Automated Structure Discovery
This study demonstrates that generative adversarial style translation effectively bridges the gap between simulated and experimental AFM images. By converting simulated images into realistic data, the framework allows DL models to accurately predict complex surface structures directly from laboratory measurements without needing labeled experimental datasets.
Future work could explore incorporating physical constraints into the training process to minimize unrealistic atomic predictions. Integrating confidence estimates for each predicted atom could also enable users to assess prediction reliability in real time. Overall, these advancements will support surface-structure analysis for advanced materials research.
Journal Reference
Huang, J., et al. (2026). Improving atomic force microscopy structure discovery via style-translation. npj Computational Materials. DOI: 10.1038/s41524-026-02243-2, https://www.nature.com/articles/s41524-026-02243-2.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.