Deep Learning for Defects Analysis and Atomic-resolution Imaging of Battery Materials

Abstract

The primary objective of this work is to utilize machine learning techniques, in particular a deep learning method, to explore structural details of various components of Li batteries. Layered oxide cathodes (e.g., NMC, LCO) have been known as the bottleneck of Li-ion batteries and multiple degradation mechanisms such as, irreversible phase transition, transition metal (e.g., Ni, Mn) migration, metal dissolution, surface decomposition and oxygen release have been realized as main contributors to the capacity decay of such cathode materials. Transition metal migration, which is regarded as the displacement of transition metal cations to Li vacancies during charge process, has been detected by advanced aberration corrected scanning transmission electron microscopy (AC-STEM). The AC-STEM z-contrast imaging allows for distinguishing transition metals from lighter species such as O and Li atoms with sub-angstrom spatial resolution. Nonetheless, this method cannot quantify the fraction of migrated transition metals to measure and compare the extent of degradation among various samples. This challenge has significantly obstructed the research and advancement of layered oxide cathodes and Li-ion batteries. The other area of interest is noble metal nanometer clusters (NCs), which play a crucial role in variety of battery types (e.g., Li-oxygen batteries). The position of atoms and the structure of NCs governs the unique size-dependent functionalities and hence their applications towards the desired electrochemical properties and mechanics. Thus, it is of imperative importance to evaluate NCs structures and atomic column heights for more profound understanding of their structure-properties relation. Employing AC-STEM and HRTEM the structure of these NCs could be evaluated. However, the determination of the atomic column heights is challenging due to uncertainty of microscope parameters and a noise in experimental data. To overcome these challenges in both of the applications of atomic resolution imaging as described above, we have developed a deep learning algorithm based upon the convolutional neural network (CNN). This work extends and improves the previously developed CNN1 by integrating predictive capabilities toward the detection of atomic column heights in complex NCs independent of noise and defocus in experimental measurements. In addition, an improved algorithm for vacancy detections in NMC cathodes is developed and applied. In both the cases (Li vacancies in NMC and atomic column height detection in NCs), the CNN is trained and labeled on 10000 HRTEM images obtained using the Atomic Simulation Environment2 and the QSTEM3 code. The artificial creation of the train data is beneficial for the training of the CNN, since it contains the absolute reliable ground truth against which labels are assigned. For atomic column heights evaluation two different methods of column heights prediction are proposed, resulting in alternative labeling of the computer-generated training data. Consequently, such a trained model is applied towards the detection of atomic column heights in experimentally measured NCs. A good prediction is obtained enabling fast and reliable analysis of variety of experimental measurements. In the case of Li vacancies identification, the model is able to find complex vacancy positions in the experimentally measured NMC layered structures allowing to analyze Li/Ni interaction. Consequently, the results of this combined spectroscopic and deep learning techniques can be used in density functional theory (DFT) to calculate the magnetic frustration in NMC structures. The developed CNN is based upon a U-Net architecture, which has some drawbacks comparing to inception networks such as GoogLeNet. Hence, we have also adapted inception type networks and compared them against a U-Net type. To summarize, in addition to improving the fidelity of the CNN model, the present study allows for a mechanistic interpretation of the origin of capacity fade of layered oxide cathodes as well as detection of atomic column heights in NCs for better structure-properties relation identification.