Electrostatic Discovery Atomic Force Microscopy

QUAM-AFM: A Free Database for Molecular Identification by Atomic Force Microscopy

“…This distance is larger than 1.5 Åthe descriptor used in ref as the height range where structural information can be retrieved from 3D structures with a collection of AFM images taken at different heightsin order to include aliphatic chains with sp 3 carbon atoms (methyl groups). In spite of the restrictions, we are still left with a huge data set of more than 685,000 molecules, significantly larger than those used in previous deep learning works in the field − ,− , and, more importantly, that spans relevant structural and compositional moieties in organic chemistry and, particularly, in the field of on-surface synthesis.…”

“…We have implemented a graphical user interface (GUI) which provides easy access to both the AFM images and the graphical descriptors of each molecule and allows a quick search by the CID number, composition, or IUPAC name. Although smaller AFM data sets have already been developed, − our proposal aims to be the definitive reference in the field, thanks to the comprehensive collection of molecular structures, data organization, and consideration of all the necessary elements to train reliable and reproducible deep learning models …”

“…Moreover, this data set can provide a robust and comprehensive reference for new artificial intelligence-based theoretical developments on scanning probe microscopy − and, in particular, in AFM. − Deep learning approaches have already been applied in AFM to the automated searching and identification of self-organized nanoparticle assemblies and in the case of high-resolution AFM images with CO-functionalized tips to obtain information on the molecular structure directly from a set of AFM images, to predict quantitative maps of the electrostatic potential, and to achieve automated chemical identification of 60 different molecular structures from both theoretical and experimental AFM images with a high accuracy . This last work clearly illustrates the huge amount of input data needed for the training of deep learning models.…”

Generalised deep-learning workflow for the prediction of hydration layers over surfaces

“…Our approach links strongly to the developing DL methods applied to data challenges in Scanning Probe Microscopy (SPM) [30][31][32][33]. In particular, the success of deep learning Convolutional Neural Networks (CNN) [34] in image recognition tasks has led to their application to the analysis of SPM images [35], especially in the context of molecular/surface feature/defect characterisation [36][37][38][39][40][41], scanning-probe characterisation [42,43], and techniques for autonomously-driven SPM [44][45][46].…”

Molecular Identification from AFM Images Using the IUPAC Nomenclature and Attribute Multimodal Recurrent Neural Networks

“…This high value means that, even when the model does not achieve a perfect prediction, it provides valuable chemical insight, leading to a correct IUPAC name of a similar molecule in the vast majority of the cases. The ability of machine learning models to provide relevant information from HR-AFM images is further supported by alternative approaches based on CNNs to predict accurate electrostatic fields 68 and on graph neural networks (GNNs) to extract molecular graphs. 69 The accuracy obtained in the extensive test with theoretical images, together with the results from few experimental examples taken from the literature, shows the potential of our deep learning approach trained with theoretical results to become a powerful tool for molecular identification from experimental HR-AFM images.…”