Atomic-Scale Phase Analysis of Layered Oxides

【Research Background】

Layered oxides are a class of cathode materials with significant applications in lithium-ion batteries. During battery operation, the complex phase transitions induced by electrochemical processes at the nanoscale remain one of the major obstacles to the development of the next generation of high energy density lithium-ion batteries. Generally, the phase degradation of layered oxides occurs through two main pathways: one is lattice shear caused by delithiation, and the other is lattice oxygen loss (along with concomitant cation mixing). The lattice oxygen loss and cation mixing in the O3 phase (initial phase) directly lead to the formation of a non-electrochemically active rock salt phase, while lattice shear leads to the formation of the detrimental O1 phase, further promoting crack nucleation, lattice oxygen loss, and the transition to rock salt. During battery operation, the O3 phase (initial phase), O1 phase, and rock salt phase intertwine and evolve at the nanoscale, a common multiphase coexistence state in layered oxides. Developing tools that can deeply analyze the complex coexistence states of the O1 phase, rock salt phase, and O3 phase is crucial for understanding the structure-performance relationships of layered oxides and guiding material optimization design.

The direct visualization and quantitative analysis of complex phases are key to understanding material phase transitions and their structure-performance relationships, while achieving “intelligent” atomic-scale phase segmentation and recognition has long been a dream of materials electron microscopists. Traditional transmission electron microscopy (TEM) electron diffraction and bright/dark field imaging typically can only identify phases at the nanoscale to microscale; in contrast, atomic resolution (scanning) transmission electron microscopy imaging can directly analyze phase transition information at the atomic scale, but still relies on manual interpretation. Against this backdrop, automatic phase segmentation (or phase recognition) technology based on atomic resolution imaging has gradually become a research hotspot at the intersection of artificial intelligence and transmission electron microscopy. In recent years, the development of artificial intelligence technology has greatly facilitated the advancement of “intelligent” TEM technology. Unfortunately, to date, a label-free atomic-scale phase segmentation technology has not yet been realized. Achieving this technology would not only automate the atomic-scale phase analysis of materials but also enable the quantification of phase analysis at the atomic scale.

【Work Introduction】

Professor Xin Huo-Lin’s research group at the University of California, Irvine (DeepEM Lab) has developed a label-free atomic-scale phase analysis technology for lithium-ion battery cathode materials based on previously developed super-resolution technology, combining deep learning models with geometric simulations of transmission electron microscopy images at the atomic scale. This research, titled “Deep-Learning Aided Atomic-Scale Phase Segmentation toward Diagnosing Complex Oxide Cathodes for Lithium-Ion Batteries“, was published in the international top journal Nano Letters. UC Irvine’s Dong Zhu and Wang Chunyang are the first authors of the paper.

【Full Analysis】

1. Workflow of Label-Free Atomic-Scale Phase Segmentation Technology

The deep learning model proposed in this study demonstrates efficient automation and quantification capabilities in the segmentation of the O1 phase, O3 phase, and rock salt phase of key lithium battery cathode materials, systematically constructing a label-free atomic-scale phase segmentation workflow. To address the challenges posed by the complexity of labeling STEM images, the authors developed an algorithm to extract and integrate the geometric features of different phases as shown in Figure 1a. The left column of Figure 1a displays the super-resolution experimental images obtained from the HAADF-STEM images of layered cathode oxides after processing by AtomSegNet (including the coordinates of each atomic column). Three different phases, namely the O3 phase, O1 phase, and rock salt phase, are identified by blue, red, and black, respectively. The column of Figure 1a shows the extracted geometric shapes of each phase, where the geometric shape composed of each atomic column and its surrounding atomic columns can serve as the basis for differentiation. To train the model, the authors developed a phase simulation algorithm to generate simulated images. The right column of Figure 1a shows an enlarged view of the simulated image composed of the three phases. Subsequently, the simulated image dataset was used to train the U-Net-based deep learning model (Figure 1b). To classify numerous ultra-small atomic columns, the authors integrated a hard attention mechanism, allowing the deep network to focus only on one pixel of each atomic column. By modifying the loss function computed for each pixel, this mechanism was incorporated into the model. Based on geometric simulation and model training, the proposed HaU-Net model successfully achieved phase segmentation on the experimental images processed by AtomSegNet. Figure 2 displays an example of a simulated image of a layered oxide from the training dataset. The method proposed in this paper involves simulating each phase point by point and line by line while setting clear generation and constraint rules. By cleverly integrating three phases with unique geometric features, this method can accurately simulate phase regions and their adjacent areas, providing an unrestricted large dataset for deep learning methods under unlabeled conditions, thus addressing the bottleneck of high data demand and labeling costs in artificial intelligence applications.

Figure 1 Workflow of label-free atomic-scale phase segmentation technology: geometric simulation, model training, and practical application

Figure 2 Simulated image of layered oxides containing multiple phases

2. Comparison of Two Hard Attention Mechanism Integration Methods

To classify the numerous tiny atomic column coordinates, this study integrated a hard attention mechanism into the neural network, guiding it to focus on relevant key areas. This mechanism encourages the model to concentrate on the atomic column areas while minimizing the impact of background areas. HaU-Net combines the semantic segmentation capabilities of U-Net with the hard attention mechanism, trained on a synthetic dataset derived from experimental STEM images. The authors explored integrating the hard attention mechanism into U-Net through two methods and compared the two approaches. Method one refines the atomic column coordinates in the input image to a single pixel point, while method two refines the loss function. Figure 3 evaluates the performance of these two methods on the synthetic image dataset, revealing that the loss function refinement method outperformed the pixel refinement method at every epoch during the training process. When training on a dataset of 40,000 simulated images, the final training results achieved an accuracy of 98.9% and an F1 score of 0.988 when redefining the back loss function for classifying each atomic column.

Figure 3 Comparison of two methods for integrating hard attention mechanisms during training

3. Comparison of the Order Parameter Method and the HaU-Net Model

The authors compared the order parameter method with the HaU-Net model and found that the phase segmentation capability of the HaU-Net model far exceeds that of traditional methods. Figures 4a and 4b respectively show the phase distribution maps obtained from processing the same charged state (delithiation state) layered oxide cathode experimental images using the two phase segmentation methods. The order parameter method is a classic approach specifically used to measure the degree of order at interfaces with symmetry breaking. The authors combined the order parameter with K-means unsupervised clustering to achieve phase segmentation. The results (Figure 4a) indicate that the order parameter method can effectively distinguish the rock salt phase from the layered structure phases (including O1 and O3 phases). However, due to the weak feature extraction capability of the order parameter method, it cannot distinguish between the structurally similar O1 and O3 phases. For instance, the boxed area and corresponding enlarged image in Figure 4a show representative regions where the O1 phase (in red) and O3 phase (in cyan) were misclassified. In contrast, the HaU-Net model significantly improved the accuracy of segmentation for all three phases (Figure 4b).

Figure 4 Comparison of traditional order parameter method and HaU-Net model

4. HaU-Net Phase Segmentation Technology Achieves Quantitative Analysis of Complex Phases

With the help of HaU-Net phase segmentation technology, the authors achieved visualization and quantitative analysis of phase distributions in layered oxide cathode materials. Figure 5 showcases the quantitative analysis of complex phases in layered oxides using HaU-Net. Figures 5a and b present the original STEM image of a charged layered oxide composed of multiple phases and the corresponding image processed by AtomSegNet. Figure 5c shows the phase distribution map obtained from the HaU-Net model, where the O3 phase, O1 phase, and rock salt phase are clearly identified. Based on atomic-level phase segmentation, the authors quantitatively analyzed the distribution of the three phases and their correlations. Figure 5e displays the proportion of each phase in the region (approximated as the number of atomic columns for each phase divided by the total number of atomic columns in the image). Through atomic neighbor analysis of each phase (Figures 5d and 5f), the authors found that the O3 phase and rock salt phase tend to coexist with their respective phases (the probabilities of neighboring phases being their own phases are 79% and 62.1%, respectively), while the O1 phase has similar chances of coexisting with all three phases (including itself) (approximately 1/3). This indicates that the formation of the O1 phase is quite random, while the formation of the rock salt phase is more localized within the O3 matrix. Furthermore, the authors also found that the probability of the O3 phase appearing in areas neighboring the rock salt phase is approximately twice that of the O1 phase, suggesting that the rock salt phase is more inclined to coexist with the O3 phase. This case analysis fully demonstrates the quantitative analysis capability of the HaU-Net model. In addition to phase segmentation, the authors also explored extending the HaU-Net model to defect (such as dislocation) segmentation. It is anticipated that the simultaneous realization of phase segmentation and defect segmentation will provide important tools for comprehensively revealing the structural evolution of layered oxides.

Figure 5 Quantitative analysis of complex phases in layered oxides achieved by HaU-Net phase segmentation technology

【Summary】

This work has developed the first artificial intelligence-assisted TEM atomic-scale phase segmentation technology, achieving automated quantitative analysis of complex phases in lithium battery layered oxides. It is anticipated that with more systematic experimental design and statistical analysis, deep learning-assisted atomic-scale phase segmentation technology will play an important role in the study of failure mechanisms in layered oxides. This technology is expected to expand to other material systems in the future, providing a powerful intelligent and quantitative tool for revealing complex phase transition behaviors in materials.

Dong Zhu#, C.Y. Wang#, P.C. Zou, R. Zhang, S.F. Wang, B.H. Song, X.Y. Yang, K.B. Low, H.L. Xin*. Deep-Learning Aided Atomic-Scale Phase Segmentation toward Diagnosing Complex Oxide Cathodes for Lithium-Ion Batteries, Nano Letters, (2023).

https://doi.org/10.1021/acs.nanolett.3c02441

【Author Introduction】

Xin Huo-Lin, Professor, Ph.D. from Cornell University. From 2013 to 2018, he established a three-dimensional in situ characterization research group at Brookhaven National Laboratory. In the summer of 2018, he transitioned to the Department of Physics at the University of California, Irvine, establishing the DeepEM Lab focused on artificial intelligence and energy materials research. Professor Xin Huo-Lin is an internationally renowned expert in the field of electron microscopy, serving as the conference chair for the Microscopy and Microanalysis 2020 annual meeting and as the vice chair in 2019. He is a member of the scientific advisory board for the NSLSII light source and a member of the proposal review committee for the Functional Nanomaterials Center at Brookhaven National Laboratory and Lawrence Berkeley National Laboratory. He received the Outstanding Early-Career Investigator Award from the Materials Research Society in 2021, the Burton Medal from the Microscopy Society of America, and the UCI Academic Senate Early-Career Faculty Award; he received the DOE Early Career Award in 2020 and the Global 30 Climate Action Hero Award. His research in characterization and clean energy has attracted attention from the government and large enterprises. Over the past four years, as the lead PI, he has received over five million dollars in funding from government and industry for his research group on green energy storage, electro/thermal catalysis, and soft matter materials. He serves as a reviewer for numerous journals, including Nature, Nat. Mater, Nat. Energy, Nat. Nanotechnol., Nat. Commun., Sci. Adv., Joule, Nano Lett., Adv. Mater., etc. His research focuses on artificial intelligence electron microscopy and deep learning, atomic-scale scanning transmission electron microscopy and related theories and techniques, high-energy electron tunneling theory, and three-dimensional reconstruction theory. In addition to theoretical and methodological research, he has conducted in-depth studies using three-dimensional electron tomography on lithium batteries, soft and hard matter interfaces, metal catalysts, and more. His research group has published over 300 articles, including 43 articles in top journals such as Science, Nature, Nat. Mater., Nat. Nanotechnol., Nat. Energy, Nat. Catal., Nat. Commun. (18 articles as corresponding author).

Wang Chunyang, Postdoctoral researcher at the University of California, Irvine, graduated with a bachelor’s degree in materials science and engineering from China University of Mining and Technology in 2014, and received his Ph.D. from the Institute of Metal Research, Chinese Academy of Sciences in 2019. During his Ph.D., he worked under Researcher Du Kui at the Shenyang National (Joint) Laboratory of Materials Science (now the Shenyang National Research Center for Materials Science) on in situ quantitative transmission electron microscopy, electron tomography three-dimensional reconstruction technology, and the deformation and phase transition of metal materials. He joined Professor Xin Huo-Lin’s research group in June 2019 for postdoctoral research. His main research directions include in situ transmission electron microscopy technology, electron tomography three-dimensional reconstruction technology, and cryo-electron microscopy technology in metals and energy materials. Specific topics include phase transition mechanisms and defect structure evolution of layered oxide cathode materials; lithium metal growth and its structure-performance relationships; lithium transport and micro-failure mechanisms of solid-state electrolytes; phase transitions, defect evolution, and structure-performance relationships in metal and ceramic materials. He has published over 50 academic papers in journals such as Nature, Nature Materials, Nature Nanotechnology, Nature Energy, Matter, PRL, Nano Letters, Advanced Materials, Nature Communications, JACS, Angewandte Chemie, Chemical Reviews, etc. As the first author (including co-first authors), he has published over 20 papers in top journals such as Nature, Nature Materials, Nature Energy, Matter (3), PRL, Nano Letters (5), Advanced Materials, and ACS Energy Letters. He received the MSA Postdoc Scholar Award in 2022 and organized the Transmission Electron Microscopy Three-Dimensional Imaging Symposium as the chair of M&M2020. Additionally, he serves as a member of the proposal review committee and user committee for the Functional Nanomaterials Center at Brookhaven National Laboratory (one of the U.S. Department of Energy user facilities).

【Recruitment】

Professor Xin Huo-Lin’s research group (DeepEMLab.com) welcomes students, postdoctoral researchers, and scholars dedicated to researching and expanding the fields of electron microscopy, polymers, batteries, and scalable production to join and visit. Interested individuals are encouraged to email their resumes to [email protected].

Polysulfide/Manganese Flow Battery: Low Cost, High Specific Energy, Scalable Production

2023-09-03