Decoding the local stoichiometry from HR-STEM images with convolutional neural networks (CNNs) trained by PRISMatic datasets
- Abstract number
- European Microscopy Congress 2020
- Corresponding Email
- [email protected]
- DHA.3 - Machine assisted acquisition and analysis of microscopy data
- Tianshu Jiang (1), Dr. Shuai Wang (1), Patrick Schramowski (1), Wolfgang Stammer (1), Prof. Dr. Lambert Alff (1), Prof. Dr. Kristian Kersting (1), Prof. Dr. rer. nat. Leopoldo Molina-Luna (1)
1. Technical University of Darmstadt
convolutional neural networks, STEM, stoichiometry
- Abstract text
Summary: High-resolution scanning transmission electron microscope (HR-STEM) images were obtained from a SrxMo1-xO3 (SMO) thin film. The local stoichiometry of the thin film was directly identified by analyzing different HR-STEM images based on convolutional neural networks (CNNs), a deep learning method. In order to train this model, the plane wave reciprocal space interpolated scattering matrix (PRISM) algorithm  was implemented in the open-source software Prismatic to simulate HR-STEM images as the training datasets.
Introduction: In recent years, STEM-related technologies have been widely applied to investigate material properties at the atomic scale, such as crystal structures, chemical compositions and defects. In order to analyse big data produced by STEM, machine learning based models are currently being implemented and developed by the microscopy community. For example, J. Aguiar et al.  recently demonstrated the classification of crystallographic structures from diffraction patterns and electron images assisted by CNNs. M. Ziatdinov et al.  showed how to correlate chemical identification and local defect transformation by combining deep learning with STEM. Nevertheless, there is still a blank left for extracting the local stoichiometry of a material directly from the HR-STEM image based on deep learning methods, which differs from traditional methods, such as electron energy loss spectroscopy (EELS). In this study, CNNs were adopted to identify the local stoichiometry from the corresponding HR-STEM image. PRISM algorithm  was applied to simulate HR-STEM images as training datasets for the CNNs. The SMO thin film region in the all-oxide varactor heterostructure device was selected as the example for this study .
Method/Materials: The oxide-based multilayer device was provided by P. Salg et al. . STEM lamellas were obtained from this multilayer device by using a Focused Ion Beam (FIB) system HR-STEM images were obtained along the  direction of the epitaxial SMO layer using a JEOL ARM-200F. Simulated HR-STEM images of the SxM1-xO layer (x from 0.45 to 0.55 with step-size of 0.1) were produced as training datasets via the software Prismatic. The CNNs were trained with randomly cropped patches of the depth-wise mean HR-STEM images from Prismatic. The local stoichiometries of the SMO layer were identified directly from experimental HR-STEM images by cross validating them via the trained CNNs.
Results and Discussion: HR-STEM images of the SxM1-xO layer, with local stoichiometries x=0.45, 0.50, 0.55, were simulated by using Prismatic. The CNNs were trained by these datasets and managed to classify these simulated HR-STEM images. These results could allow the direct identification of local stoichiometries within a single SMO layer in such an oxide-based multilayer device and thus, provide vital information towards the precise control of the epitaxy and properties of these multilayer heterostructure devices.
Conclusion: The use of the PRISM algorithm showed a reliable way to produce a large amount of simulation data for training the CNNs. With the help of CNNs, the local stoichiometry within the SMO layer could be identified at atomic scale directly from the corresponding HR-STEM image, which offers a novel way to analyse the impact of the local stoichiometry on the crystal structure and the thin film growth.
 P. Alan, C. Ophus and J. Miao, Advanced structural and chemical imaging 3 (2017), p. 15.
 J. Aguiar, M. Gong, R. Unocic, T. Tasdizen and B. Miller, Science Advances 5 (2019), p. eaaw1949.
 M. Ziatdinov, O. Dyck, A. Maksov, X. Li, X. Sang, K. Xiao, R. Unocic, R. Vasudevan, S. Jesse and S. Kalinin, ACS Nano 11 (2017), p. 12742.
 P. Salg, D. Walk, L. Zeinar, A. Radetinac, L. Molina-Luna, A. Zintler, R. Jakoby, H. Maune, P. Komissinskiy and L. Alff, APL Materials 7 (2019), p. 051107.