Improving atomic force microscopy structure discovery via style-translation

This is the source code for our paper https://arxiv.org/abs/2509.02240.

Abstract

Atomic force microscopy (AFM) is a key tool for characterising nanoscale structures, with functionalised tips now offering detailed images of the atomic structure. In parallel, AFM simulations using the particle probe model provide a cost-effective approach for rapid AFM image generation. Using state-of-the-art machine learning models and substantial simulated datasets, properties such as molecular structure, electrostatic potential, and molecular graph can be predicted from AFM images. However, transferring model performance from simulated to experimental AFM images poses challenges due to the subtle variations in real experimental data compared to the seemingly flawless simulations. In this study, we explore style translation to augment simulated images and improve the predictive performance of machine learning models in surface property analysis. We reduce the style gap between simulated and experimental AFM images and demonstrate the method's effectiveness in enhancing structure discovery models through local structural property distribution comparisons. This research presents a novel approach to improving the efficiency of machine learning models in the absence of labelled experimental data.

Overview

There are several folders in the project.

• src: source code for this study, which includes (1) preEvaluation: Data-driven approach to evaluate the performance of style translation; (2) performanceEvaluation: Performance evaluations of the structure models on the experimental AFM images based on the local structural properties; and (3) StyleTrans (submodule): Using CycleGAN framework for training the style translation model to obtain the style translation models. In (1) and (2), the included snakemake file is used to show the logic and run the code.
• data: input data files used in the project
• processed_data: intermediate files from the analysis. These files can be generated by the code in src.
• results: results of the analysis, which are used for the figures in the manuscript.
• manuscript: figures used in the manuscript.

Installation

We use conda to manage all the packages. You can create and activate it by using:

conda env create -f myenv.yml
conda activate sta

Run the code

1. Make sure to enter the sta conda environment as shown in Installation.
2. Change to the corresponding folder and check the Snakefile:

cd src/performanceEvaluation/

For each Python, there's a rule to run it in Snakefile:

...
# Show training sample
rule visualiseTrainingData:
    input:
        script = "visualiseTrainingData.py",
        data_dir = "../../data/structures/simulations/Label"
    output:
        out_dir = directory("../../results/train_data")
    shell:
        """
        python ./visualiseTrainingData.py
    """
...

3. Run it using 'snakemake --cores 1 visualiseTrainingData.

If you have any questions about the code, feel free to open an issue.

Data

The training data for the style translation, machine expert, and structure discovery models are publicly available at Zenodo: https://doi.org/10.5281/zenodo.16828078.

Citation

Cited as:

Jie Huang, Niko Oinonen, Fabio Priante, Filippo Federici Canova, Lauri Kurki, Chen Xu, and Adam S. Foster, Improving atomic force microscopy structure discovery via style-translation, arXiv:2509.02240, 2025

Or

@misc{sin2025sta,
      title={Improving atomic force microscopy structure discovery via style-translation}, 
      author={Jie Huang and Niko Oinonen and Fabio Priante and Filippo Federici Canova and Lauri Kurki and Chen Xu and Adam S. Foster},
      year={2025},
      eprint={2509.02240},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2509.02240}, 
}