Enhancing AFM image analysis through machine learning with style translation and data augmentation

Jie Huang¹, Niko Oinonen^{1, 2}, Lauri Kurki¹, Adam S. Foster^{1, 3}

¹Department of Applied Physics, Aalto University, Finland
²Nanolayers Research Computing Ltd., London N12 0HL, United Kingdom
³WPI Nano Life Science Institute (WPI-NanoLSI), Kanazawa University, Japan

Atomic Force Microscopy (AFM) is critical for atomic-scale nanostructure characterization. Simulations, especially using Particle Probe AFM (PPAFM), provide a cost-effective means for rapid image generation. Leveraging state-of-the-art machine learning models and substantial PPAFM-generated datasets, properties like molecular structures, electrostatic force potential, and molecular graphs can be accurately predicted using AFM images from simulations or experiments. However, transferring model performance from PPAFM to real AFM images poses challenges due to the subtle variations in real experimental data compared to the seemingly flawless nature of simulations. Our study explores Cycle GANs for style translation to augment data and improve the predictive accuracy of machine learning models in surface property analysis. Focused on mitigating the gap between simulated PPAFM and authentic AFM images, we optimize hyperparameters, showcasing the method's effectiveness through paired data comparisons. This research promises valuable insights, providing a novel approach to enhance machine learning model efficiency in the absence of abundant experimental data.