A real-time deep learning model to narrow the gap between atomic scanning transmission electron microscopy and theory calculations: Recognition, reconstruction, and simulation
Danpeng Cheng,Wuxin Sha,Yaqing Guo,Shun Tang,Jun Zhong,Jinqiao Du,Jie Tian,Yuan-Cheng Cao
DOI: https://doi.org/10.1557/s43577-024-00814-6
IF: 5
2024-12-12
MRS Bulletin
Abstract:The crystal structure evolution of LiNixCoyMn1−x−yO2 (NCM) lithium-ion batteries has a significant impact on battery performance. Using the crystal structure observed by advanced scanning transmission electron microscopy (STEM) equipment as input for simulation calculations, rather than theoretical atomic configurations, can improve simulation accuracy. The rise of deep learning models has enhanced the analytical efficiency of STEM images and brought dawn to battery research. However, there are still very few atomic detection models that can be generalized to different crystal structures, such as the layered phase and rock-salt phase of NCM. We propose a novel AtomSeg model that enables one-stage, real-time atomic detection in STEM images, significantly enhancing segmentation accuracy. Additionally, we introduce a framework that combines AI-driven segmentation results with simulation calculations, leveraging real atomic coordinates to more accurately reflect the actual atomic states. AtomSeg is only trained on simulated STEM images of rock-salt phases while it can be implemented in real STEM images of different scales and phases. Compared to previous deep learning methods, AtomSeg has improved the accuracy of identifying the positions and types of atoms and defects. Subsequently, AtomSeg was used to segment instances of STEM images from NCM, and the segmented localization information was used to reconstruct a three-dimensional atomic model, which was then used for simulation calculations of lithium migration properties. This research paradigm can improve the utilization of STEM data and the accuracy of simulation calculations, effectively promoting the integration of deep learning with material experimental characterization and simulation calculations. The groundbreaking advancements in understanding the crystal structure evolution of LiNixCoyMn1−x−yO2 (NCM) lithium-ion batteries, as detailed in our research, have profound implications for energy-storage devices. By leveraging advanced scanning transmission electron microscopy (STEM) technology and integrating it seamlessly with simulation calculations, we have paved a new path toward more precise and predictive battery design. The emergence of deep learning has further accelerated materials research progress, revolutionizing the way STEM images are analyzed. However, the scarcity of generalized atomic detection models for diverse crystal structures still hinders the applications of deep learning analytical tools. Our proposed AtomSeg model can identify atomic positions and types with unprecedented accuracy in real STEM images, demonstrating the feasibility of real-time atomic detection across different scales and phases. Moreover, the utilization of AtomSeg-segmented STEM images to reconstruct three-dimensional atomic models has opened up new avenues for simulating lithium migration properties. This achievement underscores the power of deep learning in bridging the gap between experimental characterization and computational modeling. In summary, this research paradigm represents a paradigm shift in battery research, demonstrating how advanced computational techniques can be harnessed to unlock the full potential of experimental data. The central innovation lies in our approach of utilizing experimentally observed crystal structures as input for simulation models, as opposed to relying solely on theoretical atomic configurations. This shift has significantly boosted the accuracy of simulations. We believe it will establish a new framework for materials research and development.
materials science, multidisciplinary,physics, applied