Abstract: We present a large dataset containing simulated core-loss electron energy loss spectroscopy (EELS) spectra with the elemental content as ground-truth labels. Additionally we present some neural networks trained on this data for element identification.  The simulated dataset contains zero padded core-loss spectra from 0 to 3072 eV, which represents 107 core-loss edges through all 80 elements from Be up to Bi. The core-loss edges are calculated from the generalised oscillator strength (GOS) database presented by Zhang et al.[1] Generic fine structures using lifetime broadened peaks are used to imitate fine structure due to solid-state effects in experimental spectra. Generic low-loss regions are used to imitate the effect of multiple scattering. Each spectrum contains at least one edge of a given query element and possibly additional edges depending on samples drawn from The Materials Project [2]. The dataset contains for each of the 80 elements: 7000 training spectra, 1500 test spectra, 600 validation spectra and 100 spectra representing only the query element. This results in a total 736 000 labeled spectra. Code on how to  – read the simulated data – transform HDF5 format to TFRecord format – train and evaluate neural networks using the simulated data – use the trained networks for automated element identification is available on GitHub at arnoannys/EELS_ID A full report on the simulation of the dataset and the training and evaluation of the neural networks can be found at:                    Annys, A., Jannis, D. & Verbeeck, J. Deep learning for automated materials characterisation in core-loss electron energy loss spectroscopy. Sci Rep 13, 13724 (2023). https://doi.org/10.1038/s41598-023-40943-7 [1] Zezhong Zhang, Ivan Lobato, Daen Jannis, Johan Verbeeck, Sandra Van Aert, & Peter Nellist. (2023). Generalised oscillator strength for core-shell electron excitation by fast electrons based on Dirac solutions (1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7729585 [2] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, Kristin A. Persson; Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Mater 1 July 2013; 1 (1): 011002. [https://doi.org/10.1063/1.4812323](https://doi.org/10.1063/1.4812323)

Abstract: Nanomaterials are materials that have at least one dimension in the nanometer length scale, which corresponds to a billionth of a meter. When three dimensions are confined to the nanometer scale, these materials are referred to as nanoparticles. These materials are of great interest since they exhibit unique physical and chemical properties that cannot be observed for bulk systems. Due to their unique and often superior properties, nanomaterials have become central in the field of electronics, catalysis, and medicine. Moreover, they are expected to be one of the most promising systems to tackle many challenges that our society is facing, such as reducing the emission of greenhouse gases and finding effective treatments for cancer. The unique properties of nanomaterials are linked to their size, shape, structure, and composition. If one is able to measure the positions of the atoms, their chemical nature, and the bonding between them, it becomes possible to predict the physicochemical properties of nanomaterials. In this manner, the development of novel nanostructures can be triggered. However, the morphology and structure of nanomaterials are highly sensitive to the conditions for relevant applications, such as elevated temperatures or intense light illumination. Furthermore, any small change in the local structure at higher temperatures or pressures may significantly modify their performance. Hence, three-dimensional (3D) characterization of nanomaterials under application-relevant conditions is important in designing them with desired functional properties for specific applications. Among different structural characterization approaches, transmission electron microscopy (TEM) is one of the most efficient and versatile tools to investigate the structure and composition of nanomaterials since it can provide atomically resolved images, which are sensitive to the local 3D structure of the investigated sample. However, TEM only provides two-dimensional (2D) images of the 3D nanoparticle, which may lead to an incomplete understanding of their structure-property relationship. The most known and powerful technique for the 3D characterization of nanomaterials is electron tomography, where the images of a nanostructured material taken from different directions are mathematically combined to retrieve its 3D structure. Although these experiments are already state-of-the-art, 3D characterization by TEM is typically performed under ultra-high vacuum conditions and at room temperature. Such conditions are unfortunately not sufficient to understand transformations during synthesis or applications of nanomaterials. This limitation can be overcome by in situ TEM where external stimuli, such as heat, gas, and liquids, can be controllably introduced inside the TEM using specialized holders. However, there are some technical limitations to successful perform 3D in situ electron tomography experiments. For example, the long acquisition time required to collect a tilt series limits this technique when one wants to observe 3D dynamic changes with atomic resolution. A solution for this problem is the estimation of the 3D structure of nanomaterials from 2D projection images acquired along a single viewing direction. For this purpose, annular dark field scanning TEM (ADF STEM) imaging mode provides a valuable tool for quantitative structural investigation of nanomaterials from single 2D images due to its thickness and mass sensitivity. For quantitative analysis, an ADF STEM image is considered as a 2D array of pixels where relative variation of pixel intensity values is proportional to the total number of atoms and the atomic number of the elements in the sample. By applying advanced statistical approaches to these images, structural information, such as the number or types of atoms, can be retrieved with high accuracy and precision. The outcome can then be used to build a 3D starting model for energy minimization by atomistic simulations, for example, molecular dynamics simulations or the Monte Carlo method. However, this methodology needs to be further evaluated for in situ experiments. This thesis is devoted to presenting robust approaches to accurately define the 3D atomic structure of nanoparticles under application-relevant conditions and understand the mechanism behind the atomic-scale dynamics in nanoparticles in response to environmental stimuli.