🔬 materials science

Improving atomic force microscopy structure discovery via style-translation

This study employs style translation to bridge the domain gap between simulated and experimental atomic force microscopy images, thereby enhancing the predictive performance of machine learning models for atomic structure discovery without requiring labeled experimental data.

Original authors: Jie Huang, Niko Oinonen, Fabio Priante, Filippo Federici Canova, Lauri Kurki, Chen Xu, Adam S. Foster

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jie Huang, Niko Oinonen, Fabio Priante, Filippo Federici Canova, Lauri Kurki, Chen Xu, Adam S. Foster

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of finding fingerprints at a crime scene, you are looking at the tiniest building blocks of matter: individual atoms. To see these invisible atoms, scientists use a super-powerful tool called an Atomic Force Microscope (AFM). Think of it like a blind person's cane, but instead of feeling a wall, it feels the surface of a material with a needle so sharp it can touch a single atom. As the needle drags across the surface, it creates a 3D map, a kind of "topographical photo" of the atomic world. This is crucial because understanding how atoms arrange themselves helps us design better medicines, faster computers, and stronger materials.

However, there is a catch. These atomic photos are incredibly hard to read. They look like fuzzy, noisy clouds rather than clear pictures of balls and sticks. To figure out what the atoms are actually doing, scientists usually have to run complex computer simulations to guess what the picture should look like, then compare that guess to the real photo. It's a slow, tedious game of "guess and check." Recently, scientists tried to use Artificial Intelligence (AI) to skip the guessing game and just read the pictures directly. But here's the problem: the AI was trained on perfect, computer-generated images that looked too clean. When the AI tried to look at real, messy photos from the microscope, it got confused and failed. It was like teaching a student to drive on a perfect, empty video game track, then dropping them into rush-hour traffic; they just didn't know how to handle the real-world chaos.

This paper tackles that exact problem. The researchers, led by Jie Huang and Adam S. Foster, asked: "Can we teach the AI to understand the messy, real world by making the perfect computer images look a little more like the real ones?" They didn't just add random static to the images; they used a clever type of AI called a "CycleGAN" (a style translator) to learn the specific "vibe" of real microscope photos. Imagine taking a pristine, high-definition photo of a forest and using a filter to make it look like a grainy, slightly blurry photo taken by a shaky hand in the rain. The AI learned exactly how to add that specific "grain" and "blur" to the perfect computer images.

Once they had these "fake-but-real-looking" images, they trained their structure-discovery AI on them. The results were promising. When they tested this new AI on actual experimental images of water molecules sitting on a gold surface, it did a much better job than the old AI trained only on perfect data. It didn't just guess more atoms; the atoms it found were arranged in a way that made physical sense, matching the rules of chemistry and physics. The team also invented a new way to check if the AI was right without needing to know the "true" answer beforehand. Instead of asking "Did you get the atoms right?", they asked, "Do the distances and angles between these atoms look like a real water molecule?" By checking these physical clues, they confirmed that their style-translated AI was producing much more reliable and realistic structures.

The paper suggests that by bridging the gap between the clean world of simulations and the messy world of real experiments, we can make AI a much more powerful partner in discovering the secrets of the nanoscale. It doesn't claim to have solved the problem forever, but it shows a clear path forward: if we want AI to understand the real world, we have to stop training it on perfect fictions and start teaching it to appreciate the beautiful messiness of reality.

Technical Summary: Improving Atomic Force Microscopy Structure Discovery via Style-Translation

Problem Definition
Atomic Force Microscopy (AFM) is a critical tool for visualizing atomic structures at the nanoscale, particularly when using functionalized tips. However, interpreting complex AFM images often requires expert intuition and is frequently validated through iterative Density Functional Theory (DFT) calculations and Probe Particle Model (PPM) simulations. While Machine Learning (ML) models have shown promise in automating the inverse problem—predicting atomic configurations directly from AFM images—their deployment is hindered by two primary challenges:

The Simulation-to-Real Domain Gap: Models trained on simulated AFM images (generated via PPM) often fail to generalize to experimental data. Simulated images are idealized, whereas experimental images contain noise, artifacts, and subtle distortions.
Scarcity of Ground Truth: Obtaining paired experimental AFM images with known ground-truth atomic structures is often unfeasible, making it difficult to train models directly on experimental data or systematically evaluate their performance on real datasets.

Methodology
To address these challenges, the authors propose a data-driven approach integrating style translation with structure prediction. The core methodology involves the following components:

CycleGAN Framework: The study employs a Cycle-Consistent Generative Adversarial Network (CycleGAN) to bridge the style gap between simulated (Domain $U$ $U$ ) and experimental (Domain $V$ $V$ ) AFM images. Unlike conventional GANs that generate data from noise, this model performs image-to-image translation using unpaired datasets.
- Generators ( $G_U, G_V$ ): $G_U$ translates simulated images to an experimental style, while $G_V$ performs the reverse.
- Loss Functions: The training objective combines adversarial loss (to distinguish real from synthetic), cycle consistency loss (to ensure $u \to \tilde{v} \to \hat{u} \approx u$ ), and identity loss (to prevent unnecessary transformation of images already resembling the target style).
Structure Discovery Models: The authors train convolutional neural networks to map 3D AFM images to atomic configurations. They compare models trained on:
1. Pure simulated data ( $F_U$ ).
2. Simulated data with handcrafted perturbations (e.g., Gaussian noise, cutouts) ( $F_{\bar{V}}$ ).
3. Simulated data translated to experimental style via CycleGAN ( $F_{\tilde{V}}$ ).
4. A hybrid of style translation and handcrafted perturbations ( $F_{V^\dagger}$ ).
Evaluation Strategy: Since ground-truth atomic structures for experimental images are unavailable, the authors introduce a novel evaluation metric based on local structural properties. Instead of comparing predictions to a known structure, they compare the statistical distributions of predicted structural features (e.g., bond lengths, angles, tetrahedral order parameters) against reference distributions derived from first-principles simulations (DFT).

Key Results

Style Gap Reduction: Quantitative evaluation using Wasserstein Distance (WD) and Fréchet Inception Distance (FID) demonstrates that the CycleGAN-based translation significantly reduces the distributional distance between simulated and experimental images. In contrast, handcrafted perturbations (like adding Gaussian noise or salt-and-pepper noise) failed to effectively bridge this gap and, in some cases, increased the distance from the experimental distribution.
Improved Prediction Performance: Models trained on style-translated data ( $F_{\tilde{V}}$ $F_{\tilde{V}}$ ) and hybrid data ( $F_{V^\dagger}$ $F_{V^{†}}$ ) showed superior performance on experimental inputs compared to the baseline model trained on pure simulations ( $F_U$ $F_{U}$ ).
- Qualitatively, the style-translated models predicted more atoms and exhibited enhanced robustness against image noise.
- Quantitatively, using the structural property distributions as a benchmark, the style-translated models achieved broader improvements across metrics such as oxygen-oxygen distance ( $d_{OO}$ ) and hydrogen bond geometry. The hybrid model ( $F_{V^\dagger}$ ) provided the most balanced and consistent performance across all evaluated structural properties.
Validation of Hypothesis: The results support the hypothesis that reducing the style gap in training data improves the generalization of structure discovery models to experimental AFM images, even in the absence of labeled experimental ground truth.

Significance and Claims
The paper claims to present a novel approach for improving the efficiency of ML models in surface property analysis when labeled experimental data is absent. By utilizing unpaired image-to-image translation, the method effectively synthesizes training data that mimics experimental conditions, thereby mitigating the simulation-to-real domain gap.

The authors emphasize that their work offers a practical pathway for closing the gap in AFM structure discovery. They explicitly note that while structural properties were used for evaluation, they were not used to guide model training during this study. The paper concludes by identifying future directions, such as integrating these structural properties as constraints during training to discourage physically implausible predictions and developing frameworks that provide confidence scores for individual atom predictions. The work does not claim to solve the inverse problem perfectly but demonstrates a significant step toward making ML-based structure discovery viable for real-world experimental data.

More like this