Beyond Predictions: An Interpretable Machine Learning Approach for Battery Performance Forecasting

doi:10.21203/rs.3.rs-4134415/v1

Download PDF

Research Article

Beyond Predictions: An Interpretable Machine Learning Approach for Battery Performance Forecasting

https://doi.org/10.21203/rs.3.rs-4134415/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Lithium-rich layered oxide (LRLO) hold great promise as cathode materials for lithium-ion batteries, but they face challenges due to their complex electrochemical behavior and structural instability. This study proposes an analysis framework using unsupervised learning via Principal Component Analysis (PCA) to improve the predictability and reliability of these materials. By applying PCA, we have identified key factors affecting their electrochemical performance and degradation mechanisms. This has enabled us to easily separate and elucidate oxygen and manganese redox reactions in the low-voltage range, thereby improving our understanding of how the evolution of these reactions affects the degradation of LRLO materials. The PCA-based approach proves to be highly effective in predicting performance and identifying degradation pathways, making a significant advance in the understanding and optimization of these cathodes. These findings represent a step forward in quantifying the mechanisms of electrode materials, which requires the development of models that integrate domain knowledge with data.

The transition to electric vehicles (EV) has accelerated the demand for battery materials that not only offer high energy density but are also cost-effective and reliable. [1] Lithium-rich layered oxide (LRLO) cathode materials are emerging as a sustainable choice due to their high energy density and the economic advantages of using primarily earth-abundant Mn. [2, 3] Theoretically, LRLOs hold the potential for higher energy densities through the unique simultaneous cation-anion redox activity, which distinguishes them from conventional cathode materials that rely primarily on transition metal redox processes.[4, 5] Despite their promising properties, LRLO materials face challenges that hinder their commercialization as their operating voltage degrades over cycles, reducing their potential for high energy output.[6, 7] This voltage fade, resulting from complex mechanisms within the material structure, offsets the benefits of high energy density and presents a significant barrier to practical application.

Decades of research on LRLO materials has focused on diagnosing the fundamental causes of voltage degradation and exploring mitigation strategies. [8, 9, 10, 11, 12, 13] It has been suggested that the migration of transition metals within the layered structure leads to a partial transformation into a disordered structure, inducing irreversible changes in oxygen redox activity and stability. [14] This revelation has stimulated numerous studies, but the structural and operational complexity of LRLO continues to hinder a full understanding and efficient exploitation. [7] The dual-phase heterogeneity of LRLOs, consisting of nanodomains with different redox chemistries (LiTMO2 and Li2MnO3)[15, 16], presents a challenge in gaining detailed mechanistic insights. This complexity makes it difficult to observe and control or quantify reactions at the nanodomain level, further complicating the pathways for improving LRLO performance.

Machine learning (ML) presents a promising opportunity to understand the complex mechanisms behind LRLO materials. Although it’s not a perfect solution, the capability of ML to detect subtle patterns and trends within extensive datasets offers a viable path for significant advancements in materials science.[17, 18] So far, ML has been applied in battery research for training models with computational data and for the effective screening of materials and molecules. [19, 20, 21] However, approaches that aim to describe battery cell behavior from atomic-scale calculations often lack direct experimental validation.[22] In contrast, empirical data-based strategies offer a viable complementary path. [23, 24] Research using ML for battery diagnostics, including state of charge (SOC) and lifetime predictions, has led to the development of predictive models that guide the formulation of rapid charging protocols and the design of practical solutions. [25, 26] Despite these advances in the application of machine learning to battery research, most existing performance prediction models fall into two main categories: models that based on an understanding of mechanisms through feature selection and engineering [27, 28, 29], and models that use neural networks and other deep learning techniques. [30] However, the scope of LRLO materials research is primarily to discover and understand mechanisms for which these existing models are poorly suited. [31]The first category of models is less effective in exploring unknown mechanisms because it requires some understanding of the mechanism to be discovered. Meanwhile, the second category, while powerful in prediction, often provides little insight into fundamental mechanistic processes due to their ”black box” nature. [32] Thus, the specific need for mechanistic exploration in the context of LRLO materials highlights a gap that is difficult to fill with conventional machine learning approaches.

This study presents an innovative method for analyzing the LRLO mechanism through the analysis of battery operating data using unsupervised machine learning and dimensionality reduction. By parameterizing the behavior of LRLO materials over a large design space with an unsupervised approach, it demonstrates a breakthrough method for unraveling the complex interactions between the structure and mechanisms of advanced battery materials. The model, derived from machine learning techniques that reduce data complexity, provides insightful information about the key physical phenomena that determine LRLO performance. This approach not only avoids the common problem of overfitting in high-dimensional data analysis, but also illuminates a path for optimizing LRLO materials to improve battery performance. By offering a fresh perspective in the ongoing discourse on battery material advancements through this innovative lens, it aims to foster further research and applications in the field of energy storage.

This study consists of four steps: performing Principal Component Analysis (PCA) from electrochemical data, analyzing the properties of the material from the PCA results, analyzing the degradation behavior, and predicting the performance from the principal components. (Fig S1) In this study, we utilized electrochemical data from previous research efforts focused on the performance of LRLO materials. Over 20,000 charge curves were analyzed, ranging from fresh to degraded states, and representing a range of physical characteristics (composition, morphology, crystallinity, etc.). To ensure a consistent study framework, we used data that held constant variables critical to material properties—including cell geometry, current density, temperature, electrode thickness, and voltage cutoff. In addition, we filtered out cycles affected by activation phenomena and lithium dendrite formation to ensure data integrity (Figure S2). Figure 1b shows the charge curves selected for model training. We chose V-I curves for their informative value when operating within a predetermined voltage range, as opposed to I-V curves, which are more commonly used in battery research.[30]

Data Generation. Our research is based on an in-house dataset of coin cells, each with cathodes composed of LRLO materials (Li_1 + xNi_yCo_zMn_{1−x−y−z}O₂, 0.05 ≤ x ≤ 0.2, 0.25 ≤ y ≤ 0.40, 0 ≤ z ≤ 0.1). These materials were synthesized with varied compositions and processing parameters for unique physical properties. Initially, the cells were cycled through initial cycles and variable rate cycles to activate essential anionic redox reactions and to assess rate capability, respectively. In addition, to ensure stability and reliability of data, a relaxation period was included every 25 cycles to facilitate cell recovery and provide detailed longitudinal performance analysis under specific conditions.

Data Preparation. After the data generation phase, which included initial and varied-rate cycles as well as relaxation periods, we engaged in a data preparation phase. This phase involved applying filters to eliminate early cycles, relaxation cycles, and any cycles that are not representative of the consistent degradation behavior typical of the materials being studied. Our goal was to isolate cycles that truly represented the performance of the material over time. We used only charge curves, because they provide more consistent and reliable information than discharge curves in real-world applications. [33] After this rigorous filtering process, our dataset comprised charge curves over 20,000 cycles, providing a robust foundation for our predictive modeling efforts. (Figure S2) We further refined the dataset by interpolating the charge curves to ensure uniform data intervals across the [3.0, 4.5]V voltage range. We then divided it into training and test subsets in an 8:2 ratio, preparing the ground for the curve prediction phase.

Feature Extraction. In this research, unsupervised learning based Principal Component Analysis (PCA) was used as a feature extraction method for charge curves. PCA was performed on a training dataset consisting of Ntrain observations to reduce the dimensionality of the high-dimensional data. This technique aims to capture the significant variance within the data by mapping it onto lower dimensional principal component axes, with further explanations of PCA available elsewhere.

The PCA procedure can be described as:

X_train ≈ XPCA = U V^T + ̄1 ̄µ^T (1)

where,

U is N_train × K -dimensional object patterns (or score patterns), which are the projections of the N_train data points into a reduced K dimensional space.
V^T is K ×N_interval -dimensional variable patterns (load patterns), which represents the contribution of each object pattern to capacity at different voltages. N_interval represents interpolated data points across the voltage range.
̄µ^T is a per-feature empirical mean, which was calculated for the dataset to normalize the data.

The object patterns serve as the reduced K dimensional coordinates for each data point, providing a simplified representation of the original high-dimensional dataset. Each column within the object pattern is representative of the weight attributed to each principal component. The variable patterns, on the other hand, detail how each dimension contributes to the capacity across voltage ranges. Each row within the variable pattern represents the variation of a principal component across voltages, and is therefore referred to as a principal component. These principal components are used to characterize redox mechanism, identify degradation mechanism of materials, and predict charge curve.

Curve Reconstruction using PCA. The principal components obtained by PCA can be interpreted as global key features representing the battery charge curve. It could be hypothesized that the original charge curve can be reconstructed by a linear combination of these principal components. In order to validate this concept, curve reconstruction was performed based on the principal components extracted from feature extraction, following the equation given in Eq. (1). After reconstruction, performance evaluation was performed by comparing each charge curve in X_train with its reconstructed counterpart obtained by the inverse transform process, X_PCA, using Root-Mean Squared Error (RMSE) as a metric.

Curve Prediction Using Features. The approach of learning the coefficients of a linear combination for segments of principal components and then applying them to the entire spectrum was considered. Accordingly, the task in this study was to predict the entire charge curve using predefined voltage intervals [v_start, v_end] for X_PCA. To obtain the coefficients corresponding to each component in the linear combination, linear regression was performed on the ground truth test dataset as described in Eq. (2).

y_ij = ̄µ_j + w_i1x_1j + w_i2x_2j + ... + w_ikx_kj (2)

Here, for 1 ≤ i ≤ N_test, v_start ≤ j ≤ v_end, and 1 ≤ k ≤ K, y_ij is the voltage corresponding to each tick location of the i^th charge curve in X_test, and x_kj represents the j^th element value corresponding to the k^th principal component in V^⊤. w_ik is the linear coefficient corresponding to x_kj associated with the k^th component.

5-fold cross-validation was used to identify the most appropriate parameters for the linear regression (LR) model. This approach allowed the prediction of residuals by excluding the per-feature empirical mean for each test charge curve. The per-feature empirical mean was then added back in to obtain the final prediction curve. The effectiveness of the LR model was evaluated using the RMSE applied over the full extent of the charge curve.

Feature extraction of LRLO charge curve using Principal Component Analysis. Principal Component Analysis (PCA), an unsupervised learning algorithm known for reducing multidimensional data into fewer dimensions while minimizing information loss, was applied to analyze a dataset of over 20,000 charge cycles from LRLO cells, each cycle with unique physical characteristics and varying states of performance degradation (Fig. 1b). By reducing correlations between components, PCA successfully identifies a set of orthogonal, or independent, features within the data. [34, 35] This approach is based on the premise that the electrochemical properties of materials are controlled by a limited number of important independent factors. [36] By applying PCA, we aimed to simplify the complex dataset into a more interpretable form by highlighting the features that most significantly explain the variance in the charge curves.

Figure 1a illustrates the process of extracting principal components and their weights from the charge curves and then reconstructing the original curves from these components. The principal components that most efficiently capture the variance in the charge curves are shown in Fig. 1c. These components, which resemble V-I curves, denote charge variations across voltage ranges, similar to the dQ/dV curves of ten used in battery research. [37, 38, 39] This similarity can help researchers intuitively understand the relation of each component in terms of redox potentials. The demonstrated independence of these components, as shown in Fig. 1d, ensures an accurate representation of the electrochemical behavior of the dataset by reducing correlations.

Using the top seven components as the shared features of LRLO cells enabled accurate reconstruction of the original charge curves from their corresponding weights, with an average error of 0.16 mAh/g (and 0.32 mAh/g for 95% of the data), as shown in Fig. 1e. (see Supplementary Section 1 and Figure S3, 4, Table S1 for detailed accuracy for the number of components used) This reconstruction accuracy demonstrates the efficiency of PCA in capturing the essential dynamics of LRLO cell cycles. In the following section, the physical significance of these principal components will be explored to further understand their influence on the electrochemical properties of LRLO cells.

Physical interpretation of principal components. A comprehensive correlation analysis was performed to elucidate the relationship between PCA-derived features and the electrochemical performance of LRLO materials. Principal components identified without prior consideration of electrochemical properties, such as capacity and average voltage, showed significant correlations with these electrochemical performance metrics. (Fig. 2a) In particular, the first principal component (PC1) showed a strong correlation with both capacity and average voltage, while the second principal component (PC2) showed a weak correlation with average voltage. (See Supplementary Section 2 for detailed relations) The sum of these two principal component curves accounts for 97% of the explanatory power, indicating that PC1 and PC2 can capture most of the characteristics of the electrochemical behavior. (Figure S5)

The correlation between principal components and electrochemical performance was found to be remarkably consistent throughout the cycling phases. (Fig. 2b, c, S6) This was further validated by detailed correlation analyses comparing materials at cycle 10 (minimal performance degradation) and cycle 50 (extensive performance degradation). (Figure S7-9) Contrary to the expectation that the factors determining the performance of initial and degraded samples might be different,[40] the correlation between PC1 and capacity and average voltage at these two cycle stages was surprisingly similar. This finding suggests a fundamental underlying physics governing both the initial performance and the degradation process of lithium-rich materials. The cycle-independent correlation between key components and key performance metrics challenges the conventional understanding of the performance degradation of lithium-ion materials. This consistency implies that the same physical principles affect both the initial electrochemical performance and the response to degradation. Such insights are valuable for LRLO material design, suggesting that optimizing materials for initial performance may inherently provide benefits with respect to degradation.

To understand the physical meaning of the principal components, we examined how different weights assigned to these components affected certain characteristics observed in our data. Specifically, we examined the relationship between the weight of the first principal component (PC1) and battery capacity, as shown in Fig. 2d. This figure shows that a lower PC1 weight is associated with an increase in charge capacity and vice versa. This correlation suggests that variations in charge capacity are associated with changes in battery behavior at intermediate voltages, particularly around 3.8V. Further analysis of the derivative of charge with respect to voltage (dQ/dV) curves, shown in Fig. 2e, allows us to identify the factors that cause shifts in the positions of the reactions that occur at intermediate voltages. We observe that an increase in the weight of PC1 not only shifts these reaction peaks to higher voltages, but also diminishes reactions at lower voltages, around 3.4 V, contributing to a decrease in overall capacity. These shifts can be indicate increased resistance within the cell. [41, 42] The increase in resistance indicated by these voltage shifts adversely affects oxygen redox reaction. This reaction, which is particularly limited by its kinetic properties, is further impaired as resistance increases, resulting in a noticeable decrease. [43, 44] The underlying factors of this increased resistance include material properties such as surface area [45, 46] and conductivity, [47] as well as degradation factors such as the formation of solid-electrolyte interphase (SEI) layer and the occurence of side reactions.[48, 49]

Charging curves with the same PC1 weight but significantly varying PC2 weights show differences in average voltage rather than capacity as illustrated in Fig. 2f. dQ/dV curves in Fig. 2g show that the PC2 weight is related to the amount of low-voltage reaction and intermediate-voltage reaction. Specifically, a lower PC2 weight results in smaller low-voltage reaction and larger intermediate-voltage reaction, while a higher PC2 weight results in the opposite. The inverse movement of low voltage reaction height and intermediate-voltage reaction height suggests their relevance to average voltage rather than capacity.

The variation of low-voltage and center reaction heights in opposite directions with changes in PC2 weight could be related to the difference in cation/anion redox ratios or Mn reduction. Material factors could include composition,[50, 51] while degradation factors could include the redox involvement of Mn due to oxygen gas evolution.[6, 9, 52, 53] Extending these findings, the first and second principal components provide a lens through which specific physical phenomena affecting battery performance can be distinguished, particularly the thermodynamic and kinetic properties associated with oxygen and manganese redox in the low-voltage region. The charging process of LRLO cells typically involves three reaction domains: intermediate-voltage reaction associated with the oxidation-reduction of Ni and Co, high-voltage reaction driven primarily by oxygen redox, and more complex low-voltage reaction involving both oxygen and Mn3+/4 + redox. [43]Separation of oxygen and Mn redox in the low-voltage reaction region usually requires careful absorption spectrum analysis. [54] The principal components derived in this study, which capture correlations between reactions in the low-voltage region and those in other voltage regions, can help isolate low-voltage reaction. The increase/decrease in low-voltage reaction by PC1 is paired with the upward/downward shift of intermediate-voltage reaction, while the increase/decrease in low-voltage reaction by PC2 is paired with the weakening/strengthening of intermediate-voltage reaction intensity. Matching these observations with known facts, the changes in low-voltage reaction caused by PC1 are associated with the deactivation of oxygen redox due to morphological degradation. In contrast, the changes in low-voltage reaction caused by PC2 are attributed to the activation of Mn redox by TM migration.

This detailed analysis confirms that the changes in size, position, and height of intermediate-voltage and low-voltage reaction affected by the weights of PC1 and PC2 play a crucial role in defining charge capacity and average voltage. This strengthens the link between the data-driven PCA model and specific physical phenomena governing battery performance, and improves our understanding of electrochemical kinetics in LRLO electrodes.

Deciphering degradation behavior of LRLO cells using PCA. Tracking the key components of the degradation process allows understanding and explanation of the degradation behavior. X-ray-based diffraction and absorption analyses are instrumental in elucidating the degradation mechanisms of electrode materials and play a crucial role in revealing the mechanisms of LRLO materials. However, tracking the continuous degradation process of LRLO materials poses significant challenges due to the complex crystal structures, domains, and composite structures of the materials, which often obscure clear analytical insights. [55, 56] To overcome these limitations, PCA was used as a model that can explain the underlying physical phenomena of electrode degradation without relying directly on analytical techniques. As shown in Fig. 3a, the PCA model reveals distinct patterns of cyclic changes in PC1 and PC2 in different cells.

Most LRLO cells exhibit similar degradation patterns, with the behavior shown in Fig. 3b serving as a representative example. This pattern, characterized by an initial increase in PC2 during step i, followed by a simultaneous increase in both PC1 and PC2 in step ii, is consistent with findings from previous studies supported by controlled experiments and advanced analyses. [57] The increase in PC2, as shown in step i of Fig. 3c, corresponds to a decrease in average voltage. This decline indicates a decreases in redox peak heights coupled with an increase in low-voltage reaction, as shown in Fig. 3d. This change suggests an increase in manganese redox activities, along with a decrease in nickel and cobalt redox activities, due to structural degradation. [6, 9, 52, 53] Furthermore, the simultaneous increase in PC1 and PC2 in step ii, suggesting an increase in redox peak positions, as shown in Fig. 3e, is indicative of an increase in material resistance. [41]

Cases where only PC2 increases (Fig. S10) and where both PC1 and PC2 increase from the beginning (Fig. S11) are also observed, which basically reflect the typical pattern, but with subtle differences that are inferred to be due to the characteristics of the material. [58] These subtle differences are speculated to complicate mechanism research. Due to incomplete data measurement and labeling for all samples, comparing mechanisms based on the physical and chemical properties of the samples is beyond the scope of the current research and remains a direction for future studies.

In addition, Figure S12 highlights a rare pattern characterized by a decrease in both PC1 and PC2 values. This pattern, associated with an increase in charge capacity, can be interpreted as activation process of anion redox. Despite efforts to exclude the initial 9 cycles involved in activation from the dataset, it appears that data reflecting prolonged activation has been included. [59] This behavior suggests that certain materials can have long-lasting activation, and such materials can be detected by PCA analysis. In addition, the changes in PC1 and PC2 values during the activation process, which are opposite to those during the degradation process, indicate that activation and degradation are electrochemically opposite phenomena.

Predictive Modeling of LRLO Charge Curves with PCA. Figure 4a illustrates a methodological framework for applying principal components to accurately predict the charge curves of unknown LRLO cells beyond the scope of the initial training dataset. The utilization of linear regression to estimate the weights of the components from the PCA model facilitates the prediction of complete charge curves based on these specific components. The accuracy of the prediction is affected by the voltage window selected for prediction, as well as the number of PCA components used. The use of up to the first 7 principal components was found to be optimal for prediction accuracy, as detailed in Table S2. This finding is consistent with the earlier assertion that the electrochemical properties of materials are governed by a limited number of significant independent factors. Using an excessive number of components leads to over-fitting of the model to the training data, thereby reducing the prediction accuracy of the model.

Figure 4b shows the variation in charge curve prediction accuracy based on the size and position of the selected voltage windows. Utilizing the full charge curve range from 3.0V to 4.5V, covering the entire 100% length, allows predictions with an average RMSE of 0.38 mAh/g. This accuracy is not only due to the extensive data input, but also emphasizes the model’s effectiveness in accurately depicting the fundamental patterns of LRLO cells.

Charge curve prediction accuracy varies significantly depending on the size and the location of the segments of the voltage range. In particular, a segment covering 40% of the voltage range from 3.3V to 3.9V is sufficient to make accurate charge curve predictions with an RMSE of less than 3 mAh/g, as shown in Fig. 4c. On the contrary, excluding data below 3.8V significantly reduces the prediction efficiency. Specifically, a 40% voltage window from 3.8V to 4.4V fails to ensure accurate charge curve predictions, likely due to the dominance of reactions occurring around 3.8V. (Figure S13)

The ability of the PCA model to generalize LRLO material behaviors is further verified by comparing the PC1-PC2 plots of the test set with those of the training set, highlighting its effectiveness in dimensional reduction for materials different from those in the training set as shown in Figs. 4d and S14. The reconstruction accuracy of the test set is very close to that of the training set. (Table S3 and Figure S15) This result highlights the model’s ability to identify outlier scenarios-data points in the test set characterized by significantly large weights in both PC1 and PC2. These outliers, associated with significantly low capacity and average voltage, exhibit phase separation in XRD analyses as shown in inset of Fig. 4d, consistent with previous reports. [60] The PCA model’s ability to simplify the complexity of electrochemical data and easily identify anomalies underscores its practical utility.

The unique adaptability of the PCA model becomes even more apparent when test datasets, which includes abnormal data, are incorporated into training. Adding data beyond the range of the existing training set does not fundamentally alter the structure of the principal components, demonstrating the model’s resilience to the incorporation of outlier data. (see Supplementary Section 3 for details, Figure S16, S17) This resilience reduces the need to retrain the model when new data are added, thus ensuring sustained high performance. [61, 62] By seamlessly integrating additional data, including instance of phase separation identified by significant PC1 and PC2 weights, the PCA model not only maintains its core framework, but also enhances its predictive accuracy and efficiency in a dynamically evolving research environment.

This research explores the relationship between the electrochemical performance and structural complexity of LRLO materials. It methodically examines their electrochemical behavior and degradation mechanisms using PCA-based unsupervised learning. By extracting physically interpretable features from experimental data, we have clearly diagnosed the performance degradation issues of LRLO materials and identified key physical phenomena that significantly affect cycling performance. In particular, this study sheds light on the decoupling of oxygen and manganese redox in the low-voltage reaction region, providing new insights that enhance our understanding of the material and point to avenues for performance improvement. Building on these insights, our analysis shows that the principal components clearly reveal specific physical phenomenon-specifically, the thermodynamic and kinetic properties associated with the oxygen and manganese redox at low voltage-that play a critical role in battery performance. This study highlights the utility of data-driven models and calls for future research to focus on integrating data with domain knowledge through interpretable models. Such integration seeks to deepen understanding of material design and operational impacts on battery aging and performance. Through this approach, we aim to forge effective strategies for enhanced battery design and material optimization, boosting electrochemical performance.

Acknowledgements

This work was supported by Korea Institute of Science and Technology Information with supercomputing resources including technical support (KSC-2022-CRE-0402).

Author contribution

J. K, I. C and I. P planned the project. J. K conducted all the electrochemical experiments and characterizations. I. C and I. P derived all the equations and programmed all the code with help from H. H. I. P supervised the project. The manuscript was written by J. K, I.C and I. P. All the authors contributed to the discussions and reviewed the manuscript.

Declaration of interests

The authors declare no competing interests.

John B. Goodenough and Kyu-Sung Park. “The Li-Ion Rechargeable Battery: A Perspective”. In: Journal of the American Chemical Society 135.4 (2013)
Gaurav Assat and Jean-Marie Tarascon. “Fundamental understanding and practical challenges of anionic redox activity in Li-ion batteries”. In: Nature Energy 3.5 (May 1, 2018), pp. 373–386.
Robert A. House, John-Joseph Marie, Miguel A. P ́erez-Osorio, Gregory J. Rees, Edouard Boivin, and Peter G. Bruce. “The role of O2 in O-redox cathodes for Li-ion batteries”. In: Nature Energy 6.8 (Aug. 1, 2021), pp. 781–789.
Zhonghua Lu, L. Y. Beaulieu, R. A. Donaberger, C. L. Thomas, and J. R. Dahn. “Synthesis, Structure, and Electrochemical Behavior of Li [ Nix Li1 / 3 2x/ 3Mn2 / 3 x / 3 ] O 2”. In: Journal of The Electrochemical Society 149.6 (Apr. 2002).
Dong-Hwa Seo, Jinhyuk Lee, Alexander Urban, Rahul Malik, ShinYoung Kang, and Gerbrand Ceder. “The structural and chemical origin of the oxygen redox activity in layered and cation-disordered Li-excess cathode materials”. In: Nature Chemistry 8.7 (July 1, 2016), pp. 692–697.
Enyuan Hu, Xiqian Yu, Ruoqian Lin, Xuanxuan Bi, Jun Lu, Seongmin Bak, Kyung-Wan Nam, Huolin L. Xin, Cherno Jaye, Daniel A. Fischer, Kahlil Amine, and Xiao-Qing Yang. “Evolution of redox couples in Li- and Mn-rich cathode materials and mitigation of voltage fade by reducing oxygen release”. In: Nature Energy 3.8 (Aug. 1, 2018), pp. 690–698.
M. Sathiya, A. M. Abakumov, D. Foix, G. Rousse, K. Ramesha, M. Sauban`ere, M. L. Doublet, H. Vezin, C. P. Laisa, A. S. Prakash, D. Gonbeau, G. Van-Tendeloo, and J-M. Tarascon. “Origin of voltage decay in high-capacity layered oxide electrodes”. In: Nature Materials 14.2 (Feb. 1, 2015), pp. 230–238.
Donggun Eum, Byunghoon Kim, Sung Joo Kim, Hyeokjun Park, Jinpeng Wu, Sung-Pyo Cho, Gabin Yoon, Myeong Hwan Lee, Sung-Kyun Jung, Wanli Yang, Won Mo Seong, Kyojin Ku, Orapa Tamwattana, Sung Kwan Park, Insang Hwang, and Kisuk Kang. “Voltage decay and redox asymmetry mitigation by reversible cation migration in lithium-rich layered oxide electrodes”. In: Nature Materials 19.4 (Apr. 1, 2020), pp. 419–427
Kyojin Ku, Jihyun Hong, Hyungsub Kim, Hyeokjun Park, Won Mo eong, Sung-Kyun Jung, Gabin Yoon, Kyu-Young Park, Haegyeom Kim, and Kisuk Kang. “Suppression of Voltage Decay through Manganese De-activation and Nickel Redox Buffering in High-Energy Layered Lithium-Rich Electrodes”. In: Advanced Energy Materials 8.21 (2018).
Jicheng Zhang, Qinghua Zhang, Deniz Wong, Nian Zhang, Guoxi Ren, Lin Gu, Christian Schulz, Lunhua He, Yang Yu, and Xiangfeng Liu. “Addressing voltage decay in Li-rich cathodes by broadening the gap between metallic and anionic bands”. In: Nature Communications 12.1 (May 24, 2021), p. 3071
Jianming Zheng, Meng Gu, Arda Genc, Jie Xiao, Pinghong Xu, Xilin Chen, Zihua Zhu, Wenbo Zhao, Lee Pullan, Chongmin Wang, and Ji-Guang Zhang. “Mitigating Voltage Fade in Cathode Materials by Improving the Atomic Level Uniformity of Elemental Distribution”. In: Nano Letters 14.5 (2014).
Ho-Young Jang, Donggun Eum, Jiung Cho, Jun Lim, Yeji Lee, Jun-Hyuk Song, Hyeokjun Park, Byunghoon Kim, Do-Hoon Kim, Sung-Pyo Cho, Sugeun Jo, Jae Hoon Heo, Sunyoung Lee, Jongwoo Lim, and Kisuk Kang. “Structurally robust lithium-rich layered oxides for high-energy and long-lasting cathodes”. In: Nature Communications 15.1 (Feb. 12, 2024), p. 1288
John-Joseph Marie, Robert A. House, Gregory J. Rees, Alex W. Robertson, Max Jenkins, Jun Chen, Stefano Agrestini, Mirian Garcia-Fernandez, Ke-Jin Zhou, and Peter G. Bruce. “Trapped O2 and the origin of voltage fade in layered Li-rich cathodes”. In: Nature Materials (Mar. 1, 2024).
Jianming Zheng, Pinghong Xu, Meng Gu, Jie Xiao, Nigel D. Browning, Pengfei Yan, Chongmin Wang, and Ji-Guang Zhang. “Structural and Chemical Evolution of Li- and Mn-Rich Layered Cathode Material”. In: Chemistry of Materials 27.4 (2015).
Tongchao Liu, Jiajie Liu, Luxi Li, Lei Yu, Jiecheng Diao, Tao Zhou, Shunning Li, Alvin Dai, Wenguang Zhao, Shenyang Xu, Yang Ren, Liguang Wang, Tianpin Wu, Rui Qi, Yinguo Xiao, Jiaxin Zheng, Wonsuk Cha, Ross Harder, Ian Robinson, Jianguo Wen, Jun Lu, Feng Pan, and Khalil Amine. “Origin of structural degradation in Li-rich layered oxide cathode”. In: Nature 606.7913 (June 1, 2022), pp. 305–312.
Jie Li, Wenting Li, Chao Zhang, Ce Han, Xinping Chen, He Zhao, Hanying Xu, Guixiao Jia, Zelin Li, Jinxing Li, Yujuan Zhang, Xin Guo, Fei Gao, Jing Liu, and Xinping Qiu. “Tuning Li2MnO3-Like Domain Size and Surface Structure Enables Highly Stabilized Li-Rich Layered Oxide Cathodes”. In: ACS Nano 17.17 (2023). pp. 16827–16839.
Pengcheng Xu, Xiaobo Ji, Minjie Li, and Wencong Lu. “Small data machine learning in materials science”. In: npj Computational Materials 9.1 (Mar. 25, 2023), p. 42.
Lauri Himanen, Amber Geurts, Adam Stuart Foster, and Patrick Rinke. “Data-Driven Materials Science: Status, Challenges,and Perspectives”. In: Advanced Science 6.21 (2019), p. 1900808.
Omar Allam, Byung Woo Cho, Ki Chul Kim, and Seung Soon Jang. “Application of DFT-based machine learning for developing molecular electrode materials in Li-ion batteries”. In: RSC Adv. 8.69 (2018). pp. 39414–39420.
Tomofumi Nakayama, Yasuhiko Igarashi, Keitaro Sodeyama, and Masato Okada. “Material search for Li-ion battery electrolytes through an exhaustive search with a Gaussian process”. In: Chemical Physics Letters 731 (2019), p. 136622.
Logan Ward, Muratahan Aykol, Ben Blaiszik, Ian Foster, Bryce Meredig, James Saal, and Santosh Suram. “Strategies for accelerating the adoption of materials informatics”. In: MRS Bulletin 43.9 (Sept. 1, 2018), pp. 683–689.
Xi Wu, Feiyu Kang, Wenhui Duan, and Jia Li. “Density functional theory calculations: A powerful tool to simulate and design high-performance energy storage and conversion materials”. In: Progress in Natural Science: Materials International 29.3 (2019), pp. 247–255.
Ehsan Samadani, Siamak Farhad, William Scott, Mehrdad Mastali, Leonardo E. Gimenez, Michael Fowler, and Roydon A. Fraser. “Empirical Modeling of Lithium-ion Batteries Based on Electrochemical Impedance Spectroscopy Tests”. In: Electrochimica Acta 160 (2015), pp. 169–177.
Madeleine Ecker, Thi Kim Dung Tran, Philipp Dechent, Stefan K ̈abitz, Alexander Warnecke, and Dirk Uwe Sauer. “Parameterization of a Physico-Chemical Model of a Lithium-Ion Battery: I. Determination of Parameters”. In: Journal of The Electrochemical Society 162.9 (June 2015). A1836.
Zhongbao Wei, Xiaofeng Yang, Yang Li, Hongwen He, Weihan Li, and Dirk Uwe Sauer. “Machine learning-based fast charging of lithium-ion battery by perceiving and regulating internal microscopic states”. In: Energy Storage Materials 56 (2023), pp. 62–75.
Peter M. Attia, Aditya Grover, Norman Jin, Kristen A. Severson, Todor M. Markov, Yang-Hung Liao, Michael H. Chen, Bryan Cheong, Nicholas Perkins, Zi Yang, Patrick K. Herring, Muratahan Aykol, Stephen J. Harris, Richard D. Braatz, Stefano Ermon, and William C. Chueh. “Closed-loop optimization of fast-charging protocols for batteries with machine learning”. In: Nature 578.7795 (Feb. 1, 2020), pp. 397–402.
Adrian Schmidt, Elvedin Ramani, Thomas Carraro, Jochen Joos, Andr ́e Weber, Marc Kamlah, and Ellen Ivers-Tiff ́ee. “Understanding Deviations between Spatially Resolved and Homogenized Cathode Models of Lithium-Ion Batteries”. In: Energy Technology 9.6 (2021). p. 2000881.
Antti Aitio and David A. Howey. “Predicting battery end of life from solar off-grid system field data using machine learning”. In: Joule 5.12 (2021), pp. 3204–3220.
Donal P. Finegan, Juner Zhu, Xuning Feng, Matt Keyser, Marcus Ulmefors, Wei Li, Martin Z. Bazant, and Samuel J. Cooper. “The Application of Data-Driven Methods and Physics-Based Learning for Improving Battery Safety”. In: Joule 5.2 (2021), pp. 316–329.
Jinpeng Tian, Rui Xiong, Weixiang Shen, Jiahuan Lu, and Xiao-Guang Yang. “Deep neural network battery charging curve prediction using 30 points collected in 10 min”. In: Joule 5.6 (2021), pp. 1521–1534.
Alexis Geslin, Bruis Van Vlijmen, Xiao Cui, Arjun Bhargava, Patrick A. Asinger, Richard D. Braatz, and William C. Chueh. “Selecting the appropriate features in battery lifetime predictions”. In: Joule 7.9 (Sept. 2023), pp. 1956–1965.
Cynthia Rudin. “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead”. In: Nature Machine Intelligence 1.5 (May 1, 2019), pp. 206–215.
Man-Fai Ng, Jin Zhao, Qingyu Yan, Gareth J. Conduit, and Zhi Wei Seh. “Predicting the state of charge and health of batteries using data-driven machine learning”. In: Nature Machine Intelligence 2.3 (Mar. 1, 2020), pp. 161–170.
Laisuo Su, Shuyan Zhang, Alan J. H. McGaughey, B. Reeja-Jayan, and Arumugam Manthiram. “Battery Charge Curve Prediction via Feature Extraction and Supervised Machine Learning”. In: Advanced Science 10.26 (2023)
Xi Chen, Luhan Ye, Yichao Wang, and Xin Li. “Beyond Expert-Level Performance Prediction for Rechargeable Batteries by Unsupervised Machine Learning”. In: Advanced Intelligent Systems 1.8 (2019). p. 1900102.
Jiangtao Hu, Qinchao Wang, Bingbin Wu, Sha Tan, Zulipiya Shadike, Yujing Bi, M. Stanley Whittingham, Jie Xiao, Xiao-Qing Yang, and Enyuan Hu. “Fundamental Linkage Between Structure, Electrochemical Properties, and Chemical Compositions of LiNi1–x–yMnxCoyO2 Cathode Materials”. In: ACS Applied Materials & Interfaces 13.2 (2021). pp. 2622–2629.
Matthieu Dubarry, Cyril Truchot, and Bor Yann Liaw. “Synthesize battery degradation modes via a diagnostic and prognostic model”. In: Journal of Power Sources 219 (2012), pp. 204–216
Bor-Rong Chen, Cody M. Walker, Sangwook Kim, M. Ross Kunz, Tanvir R. Tanim, and Eric J. Dufek. “Battery aging mode identification across NMC compositions and designs using machine learning”. In: Joule 6.12 (2022), pp. 2776–2793.
Matthieu Dubarry and David Beck. “Analysis of Synthetic Voltage vs. Capacity Datasets for Big Data Li-ion Diagnosis and Prognosis”. In: Energies 14.9 (2021).
Wen Liu, Pilgun Oh, Xien Liu, Seungjun Myeong, Woongrae Cho, and Jaephil Cho. “Countering Voltage Decay and Capacity Fading of Lithium-Rich Cathode Material at 60 °C by Hybrid Surface Protection Layers”. In: Advanced Energy Materials 5.13 (2015). p. 1500274.
Lena Spitthoff, Preben J. S. Vie, Markus Solberg Wahl, Julia Wind, and Odne Stokke Burheim. “Incremental capacity analysis (dQ/dV) as a tool for analyzing the effect of ambient temperature and mechanical clamping on degradation”. In: Journal of Electroanalytical Chemistry 944 (2023), p. 117627.
Jaeyoung Kim, Wontae Lee, Jangwhan Seok, Sangbin Park, Joon Keun Yoon, Seung-Beom Yoon, and Won-Sub Yoon. “Electrochemical profiling method for diagnosis of inhomogeneous reactions in lithium-ion batteries”. In: Cell Reports Physical Science 4.4 (2023), p. 101331.
Gaurav Assat, Dominique Foix, Charles Delacourt, Antonella Iadecola, R ́emi Dedryv`ere, and Jean-Marie Tarascon. “Fundamental interplay between anionic/cationic redox governing the kinetics and thermodynamics of lithium-rich cathodes”. In: Nature Communications 8.1 (Dec. 20, 2017), p. 2219.
Gi-Hyeok Lee, Vincent Wing-hei Lau, Wanli Yang, and Yong-Mook Kang. “Utilizing Oxygen Redox in Layered Cathode Materials from Multi-scale Perspective”. In: Advanced Energy Materials 11.27 (2021). p. 2003227.
Yijia Shao, Chaozhong Li, Luoqian Li, Jian Liu, and Shijun Liao. “Morphology-Tuned Porous Lithium-Rich Cathode Materials Synthesized via a Solvothermal Approach for Li-Ion Battery Application”. In: ACS Applied Energy Materials 6.4 (2023). pp. 2531–2540.
Bao Qiu, Chong Yin, Yonggao Xia, and Zhaoping Liu. “Synthesis of Three-Dimensional Nanoporous Li-Rich Layered Cathode Oxides for High Volumetric and Power Energy Density Lithium-Ion Batteries”. In: ACS Applied Materials & Interfaces 9.4 (2017). pp. 3661–3666.
Zizhen Zhou, Dewei Chu, Bo Gao, Toshiyuki Momma, Yoshitaka Tateyama, and Claudio Cazorla. “Tuning the Electronic, Ion Transport, and Stability Properties of Li-rich Manganese-based Oxide Materials with Oxide Perovskite Coatings: A First-Principles Computational Study”. In: ACS Applied Materials & Interfaces
14.32 (2022). pp. 37009–37018.
Shi Tan, Zhongru Zhang, Yixiao Li, Yun Li, Jianming Zheng, Zhibin Zhou, and Yong Yang. “Tris(hexafluoro-iso-propyl)phosphate as an SEI-Forming Additive on Improving the Electrochemical Performance of the Li[Li0.2Mn0.56Ni0.16Co0.08]O2 Cathode Material”. In: Journal of The Electrochemical Society 160.2 (Feb. 2012). A285.
Feixiang Ding, Jianling Li, Fuhai Deng, Guofeng Xu, Yanying Liu, Kai Yang, and Feiyu Kang. “Surface Heterostructure Induced by PrPO4 Modification in Li1.2[Mn0.54Ni0.13Co0.13]O2 Cathode Material for High-Performance Lithium-Ion Batteries with Mitigating Voltage Decay”. In: ACS Applied Materials & Interfaces 9.33 (2017). pp. 27936–27945.
Chong Yin, Zhining Wei, Minghao Zhang, Bao Qiu, Yuhuan Zhou, Yinguo Xiao, Dong Zhou, Liang Yun, Cheng Li, Qingwen Gu, Wen Wen, Xiao Li, Xiaohui Wen, Zhepu Shi, Lunhua He, Ying Shirley Meng, and Zhaoping Liu. “Structural insights into composition design of Li-rich layered cathode materials for high-energy rechargeable battery”. In: Materials Today 51 (2021), pp. 15–26.
Edouard Boivin, Niccolo Guerrini, Robert A. House, Juan G. Lozano, Liyu Jin, Gregory J. Rees, James W. Somerville, Christian Kuss, Matthew R. Roberts, and Peter G. Bruce. “The Role of Ni and Co in Suppressing O-Loss in Li-Rich Layered Cathodes”. In: Advanced Functional Materials 31.2 (2021). p. 2003660.
Weibo Hua, Suning Wang, Michael Knapp, Steven J. Leake, Anatoliy Senyshyn, Carsten Richter, Murat Yavuz, Joachim R. Binder, Clare P. Grey, Helmut Ehrenberg, Sylvio Indris, and Bj ̈orn Schwarz. “Structural insights into the formation and voltage degradation of lithium- and manganese-rich layered oxides”. In: Nature Communications 10.1 (Nov. 26, 2019), p. 5365.
Peter M. Csernica, Samanbir S. Kalirai, William E. Gent, Kipil Lim, Young-Sang Yu, Yunzhi Liu, Sung-Jin Ahn, Emma Kaeli, Xin Xu, Kevin H. Stone, Ann F. Marshall, Robert Sinclair, David A. Shapiro, Michael F. Toney, and William C. Chueh. “Persistent and partially mobile oxygen vacancies in Li-rich layered oxides”. In: Nature Energy 6.6 (June 1, 2021), pp. 642–652.
William E. Gent, Kipil Lim, Yufeng Liang, Qinghao Li, Taylor Barnes, Sung-Jin Ahn, Kevin H. Stone, Mitchell McIntire, Jihyun Hong, Jay Hyok Song, Yiyang Li, Apurva Mehta, Stefano Ermon, Tolek Tyliszczak, David Kilcoyne, David Vine, Jin-Hwan Park, Seok-Kwang Doo, Michael F. Toney, Wanli Yang, David Prendergast, and William C. Chueh. “Coupling between oxygen redox and cation migration explains unusual electrochemistry in lithium-rich layered oxides”. In: Nature Communications 8.1 (Dec. 12, 2017), p. 2091.
Qingyuan Li, De Ning, Deniz Wong, Ke An, Yuxin Tang, Dong Zhou, G ̈otz Schuck, Zhenhua Chen, Nian Zhang, and Xiangfeng Liu. “Improving the oxygen redox reversibility of Li-rich battery cathode materials via Coulombic repulsive interactions strategy”. In: Nature Communications 13.1 (Mar. 2, 2022), p. 1123.
Wei Yin, Alexis Grimaud, Gwenaelle Rousse, Artem M. Abakumov, Anatoliy Senyshyn, Leiting Zhang, Sigita Trabesinger, Antonella Iadecola, Dominique Foix, Domitille Giaume, and Jean-Marie Tarascon. “Structural evolution at the oxidative and reductive limits in the first electrochemical cycle of Li1.2Ni0.13Mn0.54Co0.13O2”. In: Nature Communications 11.1 (Mar. 6, 2020)
Gaurav Assat, Antonella Iadecola, Dominique Foix, R ́emi Dedryv`ere, and Jean-Marie Tarascon. “Direct Quantification of Anionic Redox over Long Cycling of Li-Rich NMC via Hard X-ray Photoemission Spectroscopy”. In: ACS Energy Letters 3.11 (2018). pp. 2721–2728.
Biao Li, Zengqing Zhuo, Leiting Zhang, Antonella Iadecola, Xu Gao, Jinghua Guo, Wanli Yang, Anatolii V. Morozov, Artem M. Abakumov, and Jean-Marie Tarascon. “Decoupling the roles of Ni and Co in anionic redox activity of Li-rich NMC cathodes”. In: Nature Materials 22.11 (Nov. 1, 2023), pp. 1370–1379.
Delai Ye, Kiyoshi Ozawa, Bei Wang, Denisa Hulicova-Jurcakova, Jin Zou, Chenghua Sun, and Lianzhou Wang. “Capacity-controllable Li-rich cathode materials for lithium-ion batteries”. In: Nano Energy 6 (2014), pp. 92–102.
Junghwa Lee, Nicolas Dupre, Mihee Jeong, ShinYoung Kang, Maxim Avdeev, Yue Gong, Lin Gu, Won-Sub Yoon, and Byoungwoo Kang. “Fully Exploited Oxygen Redox Reaction by the Inter-Diffused Cations in Co-Free Li-Rich Materials for High Performance Li-Ion Batteries”. In: Advanced Science 7.17 (2020). p. 2001658.
Fan Fan, Gang Wu, Yining Yang, Fu Liu, Yuli Qian, Qingmiao Yu, Hongqiang Ren, and Jinju Geng. “A Graph Neural Network Model with a Transparent Decision-Making Process Defines the Applicability Domain for Environmental Estrogen Screening”. In: Environmental Science & Technology 57.46 (2023). pp. 18236–18245.
Christopher Sutton, Mario Boley, Luca M. Ghiringhelli, Matthias Rupp, Jilles Vreeken, and Matthias Scheffler. “Identifying domains of applicability of machine learning models for materials science”. In: Nature Communications 11.1 (Sept. 4,2020), p. 4428.

The authors declare no competing interests.

240402Supplementary.docx
Supplementary Materials

Download PDF

Version 1

posted

You are reading this latest preprint version

Beyond Predictions: An Interpretable Machine Learning Approach for Battery Performance Forecasting

Status:

Version 1

Abstract

Figures

Introduction

Experimental Section

Results and Discussion

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1