Unpaired Learning-Enabled Nanotube Identification from AFM Images

Nanotubes, particularly single-walled carbon nanotubes (SWCNTs), represent an important class of materials with valuable electrical, mechanical, and thermal properties. However, accurate characterization of nanotube network morphologies remains a significant challenge. We present a deep learning-based approach for extracting nanotube morphologies from atomic force microscopy (AFM) images utilizing an image-to-image (I2I) translation framework based on cycleGAN with a specialized loss function that learns to transform AFM images containing nanotubes to images of pure substrates. By subtracting these translated substrate images from the original AFM images, we effectively isolated nanotube morphologies even on substrates with roughness exceeding the nanotube diameter. We validate our approach through physics-based simulation studies and application to roll-to-roll printed carbon nanotube transistors on flexible polymeric substrates.Our method outperforms tranditional image processing and supervised learning models in sensitivity and accuracy for CNT network extraction. This improved characterization capability provides useful insights for optimizing the fabrication processes of CNT-TFTs, supporting their development for flexible electronic applications. The methodology extends beyond carbon nanotubes to other nanomaterial-based electronic devices, with future work aimed at expanding the model's generalization and integrating with real-time AFM imaging.