A Deep Learning Approach for Molecular Classification Based on AFM Images

In spite of the unprecedented resolution provided by non-contact atomic force microscopy (AFM) with CO-functionalized and advances in the interpretation of the observed contrast, the unambiguous identification of molecular systems solely based on AFM images, without any prior information, remains an open problem. This work presents a first step towards the automatic classification of AFM experimental images by a deep learning model trained essentially with a theoretically generated dataset. We analyze the limitations of two standard models for pattern recognition when applied to AFM image classification and develop a model with the optimal depth to provide accurate results and to retain the ability to generalize. We show that a variational autoencoder (VAE) provides a very efficient way to incorporate, from very few experimental images, characteristic features into the training set that assure a high accuracy in the classification of both theoretical and experimental images.

Automated summary

Despite the resolution provided by AFM with CO-functionalized, identifying molecular systems based solely on AFM images remains challenging. This study presents a deep learning model trained on a theoretically generated dataset for automatic classification of AFM experimental images. The study explores the limitations of standard pattern recognition models and proposes a model with optimal depth for accurate results and generalization ability. A variational autoencoder is shown to efficiently incorporate characteristic features from very few experimental images into the training set, ensuring high accuracy in classification.