NCAFM2023 Programme Booklet

Wednesday 1620 - 1640

MOLECULAR IDENTIFICATION FROM HR-AFM IMAGES BY PREDICTING CHEMICAL FINGERPRINTS WITH DEEP LEARNING

Pablo Pou 1,2 , Manuel Eduardo González Lastre 1 , Jaime Carracedo-Cosme 1,3 , Carlos Romero-Muñiz 4 , Rubén Pérez 1,2 * (Co-authors)

1 Departamento de Física Teórica de la Materia Condensada, Universidad Autónoma de Madrid, E-28049 Madrid, Spain 2 Condensed Matter Physics Center (IFIMAC), Universidad Autónoma de Madrid, E-28049 Madrid, Spain 3 Quasar Science Resources S.L., Camino de las Ceudas 2, E-28232 Las Rozas, Spain 4 Departamento de Física de la Materia Condensada, Universidad de Sevilla, P.O. Box 1065, 41080 Sevilla, Spain

Email: pablo.pou@uam.es

One of the fundamental challenges for the AFM is the possibility to perform chemical identification of molecules. This has been achieved but by combining HR-AFM with other techniques as STM, KPFM or simulations [1]. Only a few works explored the possibility of finding distinctive features in images or 3D force maps that could be used as a fingerprint of the chemical species of the atoms [2]. Based on these works arose the hypothesis that a stack of HR-AFM images holds chemical information of the molecules but that this is very subtle and depends not only on the chemical species of the atoms but also on their neighborhood, the molecular moieties. Machine Learning techniques are perfectly suited to extract this information. Recently several woks have used Deep learning algorithms to characterize HR-AFM [3,4,5], showing that it is possible to extract chemical information from the HR-AFM. Here we will show two ways we used to identify molecules through a HR-AFM image stack: (i) by formulating the chemical identification as an imaging captioning problem, whereby we exploit multimodal recurrent neural networks (M-RNN) to produce the IUPAC name of the imaged problem [4]; and (ii) by extracting topological chemical fingerprints of the molecules, as the Extended Connectivity Fingerprints (ECFPs), with a convolutional neural network [5]. Both networks are trained with simulated HR-AFM image stacks for ~700,000 molecules, included in the QUAM-AFM data set, and provide very good accuracy. They even produce promising results with experimental images.

Fig. Scheme of a Deep Learning algorithm that writes the IUPAC name of the imaged molecule [4]

References [1] L. Gross et al., Angew Chem 2018, 57, 3888 [2] N. J. van der Heijden et al., ACS Nano 2017, 10, 8517; M. Ellner et al., ACS Nano 2019, 13, 786 [3] B. Alldritt et al, Sci. Adv., 2020, 6, 6913; N. Oinonen et al., ACS Nano 2022, 16, 89; MRS Bull 2022, 47, 895

[4] J. Carracedo-Cosme et al, ACS AMI, 2023, 15, 22692 [5] M. E. Gonzalez Lastre et al, in preparation 2023

64