Cross-modal Deep Learning for Predicting Atomic Force Microscopy From Optical Microscope Images

In the experimental section, we demonstrate a unique approach to predicting atomic force microscopy (AF-Mic) topography from optical microscope (OptM) images by employing a fully convolutional, multi-domain autoencoder model. Topography prediction, morphology analysis, and dynamic topography transformation simulation are all made possible by the autoencoder's ability to modify input data by extracting significant features and expressing them in an enhanced fashion. The model is detailed with encoder and decoder components, optimised for various OptM image resolutions. There are three main parts to the training process: predicting the topography using the AF-Mic, analysing the morphology quantitatively, and simulating the topography's dynamic changes. Improved performance is shown after transfer learning is put into practise. When making graphene oxide nanoribbons (GON), the autoencoder is used to foretell how the material's surface would change throughout production. Accurate topographical prediction, categorization based on morphological traits, and modelling of topography transformation are only a few of the strengths of this technology, which together constitute a robust framework for grasping and keeping tabs on dynamic surface changes. Results show promise for the proposed autoencoder-based simulation to be useful in domains outside those using GON structures, while also highlighting the simulation method's relative ease of use and cost-effectiveness when compared to more conventional approaches.