Machine Learning-based Comprehensive Survey on Lithium-rich Cathode Materials

Abstract

The practical application of Li-rich cathode materials exhibiting higher energy density with oxygen redox activity requires improved cycle performance and energy efficiency. Since several conditions such as the amount of excess lithium, transition metal species, and cutoff voltage influence the electrochemical properties of Li-rich cathode materials, comprehensive determination of the optimal conditions often rely on repeating empirical try error processes. Here, the dominant factors in the energy density of Li-rich cathode materials were analyzed by constructing machine learning prediction models based on well-controlled experimental data for simplicity. Choosing a moderate amount of excess lithium and increasing the cobalt contents are the keys to achieve high energy density in long-term cycles.

1. Introduction

Lithium-ion batteries (LIBs) are to be widely deployed in a sustainable society, as represented by power sources for electric vehicles.^1,2 However, their energy density is limited in part by the small capacity of cathode materials: conventional cathode materials LiMO₂ (M = transition metal) exhibit the charge compensation by the redox reaction of transition metals in association with reversible lithium (de)intercalation. On the other hand, lithium-rich transition metal oxides Li_1+xM_1−xO₂ deliver a large capacity of >250 mAh/g using extra oxygen-redox reaction, which leads to significant increase in the energy density.^3,4 Extra oxygen-redox reaction in Li_1+xM_1−xO₂ occurs at nonbonding oxygen 2p orbital in a specific Li–O–Li bond configuration.^5–7 This principle has stimulated intensive exploration of several oxygen-redox cathodes such as Li_1.2Ni_0.13Mn_0.54Co_0.13O₂,³ Li₂Ru_0.75Sn_0.25O₃,⁴ Li_1.3Nb_0.3Mn_0.4O₂,⁸ Na₂RuO₃,⁹ and Na₂Mn₃O₇.¹⁰

However, several technical issues upon oxygen redox have been identified such as voltage hysteresis, voltage and capacity decay,^11–13 caused by oxygen dimer formation,^14,15 oxygen gas generation,^16,17 and transition metal migration.^18–21 Considering these kinetically competing side reactions, performance optimization needs simultaneous control of the synthesis conditions, charge/discharge protocols, chemical composition of Li_1+xM_1−xO₂, crystallographic polymorphs, C-rate, cut-off voltage, electrolyte, separator, electrode thickness, conducting carbon, binder, etc. While researchers have attempted to reveal the relationship between the electrochemical property and these intrinsic/extrinsic conditions,^{15,18,22–25} dominant factors in the performance of oxygen-redox cathodes are yet to be comprehensively understood.

Machine learning is a powerful methodology for analysis of critical factors and optimum conditions.^26–32 In this work, we attempted to clarify important features dominating the electrode performance of the most widely studied oxygen-redox cathodes in the Li[Li_1/3Mn_2/3]O₂–Li[Ni_1/2Mn_1/2]O₂–LiCoO₂ ternary system (Fig. 1).

Figure 1.

Overview of the machine learning approach to optimized Li-rich cathodes. (a) The Li[Li_1/3Mn_2/3]O₂–Li[Ni_1/2Mn_1/2]O₂–LiCoO₂ (Li_1+xMn_yNi_zCo_{1−x−y−z}O₂) ternary system, where each transition metal have a unique valence state (Ni²⁺, Mn⁴⁺, Co³⁺) (red: major compositions explored in previous literatures. black: compositions synthesized in this research.) (b) A schematic illustration of the research landscape.

2. Methods

2.1 Materials synthesis

Solid solutions in the Li[Li_1/3Mn_2/3]O₂–Li[Ni_1/2Mn_1/2]O₂–LiCoO₂ ternary system was synthesized by sintering coprecipitated carbonate precursors.²² The stoichiometric amount of transition metal sulphates and Li₂CO₃ were separately dissolved in 100 ml of distilled water to prepare 0.5 mol/L aqueous solution. These solutions were added dropwise to 300 mL of water at 50 °C in 1 h (about 1 drop/s) and then stirred for 1 h. The precipitate was separated from aqueous solution and dried at 180 °C for 12 h. The stoichiometric mixture of the obtained transition metal carbonate and LiCO₃ (5 % excess) was ball-milled for 4 h at 400 rpm (5 min milling − 2 min waiting), pressed into pellets and then heated at 480 °C (5 °C/min heating ramp) for 3 h under air flow. The pellets were ground and pressed again, followed by heating at 900 °C (5 °C/min heating ramp) for 6 h under air flow. The X-ray diffraction (XRD) patterns for synthesized samples were measured using RINT TTR III equipped with a CuKα radiation source.

2.2 Electrochemistry

Electrochemical measurements were carried out using a CR2032 type coin cell. Positive electrodes consisted of a mixture of 80 weight(wt)% of active materials, 10 wt% of acetylene black, and 10 wt% of polyvinylidene difluoride (PVDF) that was coated on an Al foil using N-methylpyrrolidone (NMP) as a solvent using doctor blade with 100 µm thickness (approximately 2.5 mg active material/cm²). Lithium foil was used as a negative electrode, the thickness of which is 0.6 mm and much thicker than the stoichiometric amount (∼0.01 mm). The positive and negative electrodes were separated by a glassfiber separator soaked with 150 µL of 1 mol/L LiPF₆ ethylene carbonate (EC) : dimethyl carbonate (DMC) (1 : 1 v/v%) as an electrolyte. The cells were cycled at a C/n rate (1 C = 250 mA/g, n = 1, 5, 10, 15, 20, 25) using a TOSCAT-3100 battery tester (Toyo system).

2.3 Machine learning

To construct a prediction model, electrochemical data were analyzed using random forest regressor in scikit-learn package on Python 3.^33,34 The amount of excess Li, Co/M ratio, and the charge/discharge conditions (C rate, voltage cutoff, and thickness of electrodes) were used as explanatory variables. To improve the prediction accuracy, hyper parameters were optimized by 5 × 5 grid search. The importance hierarchy was calculated by out-of-bag algorithm.³⁴

3. Results and Discussion

The XRD patterns for the lithium-rich samples synthesized in various compositions in the Li[Li_1/3Mn_2/3]O₂–Li[Ni_1/2Mn_1/2]O₂–LiCoO₂ ternary system show that most peaks except in-plane ordered superstructure peaks are attributable to a space group $R\bar{3}m$ (Fig. 2a), and the lattice dimensions linearly change as a function of the composition, both of which indicate the successful synthesis of the target samples (Fig. 2b). The charge/discharge experiments of these synthesized compounds were performed under various conditions (upper cutoff voltages, C-rates, and electrode weights) (Fig. 3, Table S1).

Figure 2.

Synthesis of Li-rich cathode materials. (a) The XRD patterns for all sample synthesized in the composition Li_1+x[Ni, Mn, Co]_1−xO₂. (b) The change in lattice parameter of c axis for Li_1+x[Ni, Mn, Co]_1−xO₂ solid solution when the amount of excess lithium is changed.

Figure 3.

Electrochemical dataset for machine learning. The schematic illustration for data collection of charge/discharge curves under various conditions. Each electrochemical property (energy density, capacity, average voltage, hysteresis, and coulombic efficiency) is extracted from these charge/discharge curves typically presented at the bottom part.

First, a machine-learning model was built from the collected data to predict energy density in the second cycle, as the energy density is the most important target for lithium-rich cathode materials. To better understand the influence of each selected feature on other electrode properties, we also built models targeted for capacity, voltage, coulombic efficiency, and voltage hysteresis. The features selected for the machine learning are the amount of excess Li, Co/M ratio, C rate, upper cutoff voltage, and electrode weight. Since the crystal structure, synthesis condition, and valence state of transition metal (Mn⁴⁺, Co³⁺ and Ni²⁺) are identical (Fig. 1) for all the compounds, intrinsic properties such as lithium-ion conductivity, electronic conductivity, redox activity, structural integrity, and particle size may correlate with the chemical composition. Therefore, the model was simplified so that only compositional information and charge-discharge conditions are learned as features, quantifying their relative importance. Note that the Co/M ratio and the amount of excess Li uniquely determine the chemical composition in the ternary system (Fig. 1).

We employed random-forest regression to predict the energy density in the 2nd cycle (Fig. 4a), because linear regression could not give prediction accuracy of satisfactory. The prediction accuracy based on the random forest regression to the test data is high enough (R² = 0.975), and the amount of excess Li is by far the most important feature in this prediction model (Fig. 4b). Note that the particle morphology, which is usually important for electrochemical properties, is included in the compositional information due to the controlled synthesis conditions. The Sharpley additive explanations (SHAP) value for each feature (Fig. 4c), that is, the impact of each feature on the prediction,³⁵ visualizes the importance of the amount of excess Li for obtaining high energy density. It is clearly visualized that low C rate, high Co/M ratio, and high cutoff voltage improve the energy density. For excess Li, which has the greatest impact, the trend is complicated in contrast to other features.

Figure 4.

The important factors for the energy density in the 2nd cycle. (a) Diagnostic plots and (b) feature importance of the prediction model for energy density of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 2nd cycle. (c) SHAP values of each feature in this prediction model.

Breaking down the energy density into capacity and voltage provides clearer view of the abovementioned strong dependency of the energy density on the amount of excess Li. While the prediction models for a capacity and an average voltage have good prediction accuracy (R² = 0.974 and 0.873) (Fig. S1), the amount of excess Li is the most influential for both target valuables of capacity and voltage (Figs. 5a and 5c). The SHAP values imply that the voltage simply decreases as the amount of excess Li increases (Fig. 5d),^15,36,37 while the moderate amount of excess Li maximizes a capacity (Fig. 5b). This trend in capacity is consistent with the previous reports that too much Li–O–Li configurations destabilize the structure and promote oxygen gas generation, resulting in significant irreversible capacity loss.^38,39 In addition, the trend in average voltage indicates that the effect of increased voltage hysteresis caused by the generation of oxygen dimers^14,15 is more significant than the effect of increased charging voltage due to the use of oxygen redox. In fact, the SHAP value for excess Li in the prediction model of voltage hysteresis shows that excess Li remarkable increase voltage hysteresis (Fig. S3). Overall, the impact of each feature on energy density is consistent with previous reports. For example, the positive contribution of Co/M ratio on energy density is consistent with the higher energy density of Li_1.2Ni_0.13Mn_0.54Co_0.13O₂ than that of Li_1.2Ni_0.2Mn_0.6O₂.⁴⁰

Figure 5.

Closer look at the energy density in the 2nd cycle by setting capacity and voltage as targets. (a) Feature importance and (b) SHAP values of the prediction model for capacity of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 2nd cycle. (c) Feature importance and (d) SHAP values of the prediction model for average voltage of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 2nd cycle.

In order to identify the key factors for achieving high energy density by stabilizing oxygen redox in long-term cycles, a random forest regression analysis was also applied to the energy density in the 50th cycle (R² = 0.827) and compared with that obtained for the second cycle (Fig. 6). Although the amount of excess Li is still the most important in the 50th cycle, its importance decreases relative to that in the 2nd cycle (Fig. 6b), and instead the influence of C rate, the electrode weight and Co/M ratio increase.

Figure 6.

The important factors for the energy density decay in long-term cycle. (a) Diagnostic plots and (b) feature importance of the prediction model for energy density of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 50th cycle. The difference in feature importance is also shown. (Difference of the feature importance in 50th cycle minus the feature importance in 2nd cycle.) (c) SHAP values of each feature in this prediction model.

The ideal Co/M ratio for stabilizing oxygen redox in long-term cycles is still under debate. According to previous research,⁴⁰ less Co/M ratio and more Ni/M ratio inhibits oxygen evolution whereas another research indicates more Co stabilizes the structure and prevents transition to spinel.⁴¹ From Fig. 6c, higher Co/M ratio is shown to enhance energy density, which is better explained by the latter research. To take a closer look at the relationship between energy density and Co/M ratio, the prediction models for capacity and average voltage in the 50th discharge are constructed and analyzed (Fig. 7). The importance of the Co/M ratio is relatively high in the prediction of capacity (∼0.15), while it remains small for voltage (∼0.04). Assuming that high Co/M ratio prevents the transition to spinel phase, Co/M is anticipated to have a greater impact on capacity than average voltage because much lower capacity of spinel phase (∼110 mAh/g) than Li-rich phase (>250 mAh/g). In fact, high Co/M ratio has a greater and clearer positive impact in capacity than average voltage according to features importance and SHAP values (Fig. 7b). Additionally, a comparison of the importance of the features at the 50th and 2nd cycles reveals that Co/M ratio has a greater impact on capacity in the 50th cycle, while its influence on average voltage remains relatively constant (Fig. S5). When conditions other than composition were fixed, the best composition was Li_1.16Ni_0.14Mn_0.49Co_0.20O₂ according to the prediction model (Fig. S7). Li-rich cathodes with compositions which contain more cobalt and less excess lithium than the conventional Li-rich cathode materials (Li_1.2Ni_0.2Mn_0.6O₂ and Li_1.2Ni_0.13Mn_0.54Co_0.13O₂) are expected to achieve better energy density in long-term cycle.

Figure 7.

Closer look at the energy density in long-term cycle by setting capacity and voltage as targets. (a) Feature importance and (b) SHAP values of the prediction model for capacity of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 50th cycle. (c) Feature importance and (d) SHAP values of the prediction model for average voltage of Li_1+x[Ni, Mn, Co]_1−xO₂ in the 50th cycle.

4. Conclusions

In summary, the impacts of chemical composition and charge/discharge conditions on the electrochemical properties of the solid solution lithium-rich transition metal oxides in the Li[Li_1/3Mn_2/3]O₂–Li[Ni_1/2Mn_1/2]O₂–LiCoO₂ ternary system were analyzed using machine learning protocols. Careful collection and selection of experimental data set have led to the reasonable prediction of the electrode performances such as energy density, capacity, and voltage as well as their degradation upon cycling. The uniformity of experimental conditions such as synthetic procedure and composition of cathode mixture simplified the analysis, resulting that the prediction models with high accuracy is successfully constructed from only compositional information and charge/discharge conditions. The derived prediction model suggested that excess Li is the most important in the energy density, where a moderate amount of excess Li is the key to maximize it. Meanwhile, increasing the Co/M ratio is the priority action to realize high energy density in long-term cycles. Although the size of data set is not large enough, a proof of concept provided in the present study may offer an opportunity to try machine learning methodology to design and optimize a battery electrode and its charge/discharge conditions, and possibly to identify the hidden factor that influences several important properties.

Acknowledgments

This work was supported by the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) Programme: Data Creation and Utilisation Type Materials Research and Development Project (Grant Number JPMXP1121467561). A. T. was financially supported by JPSP KAKENHI (Grant Number: 20J23089) and the ANRI fellowship.

Data Availability Statement

The data that support the findings of this study are openly available under the terms of the designated Creative Commons License in J-STAGE Data at https://doi.org/10.50892/data.electrochemistry.22180384.

• 1) The practical application of Li-rich cathode materials exhibiting higher energy density with oxygen redox activity requires improved cycle performance and energy efficiency. Since several conditions such as the amount of excess lithium, transition metal species, and cutoff voltage influence the electrochemical properties of Li-rich cathode materials, comprehensive determination of the optimal conditions often rely on repeating empirical try error processes. Here, the dominant factors in the energy density of Li-rich cathode materials were analyzed by constructing machine learning prediction models based on well-controlled experimental data for simplicity. Choosing a moderate amount of excess lithium and increasing the cobalt contents are the keys to achieve high energy density in long-term cycles.

CRediT Authorship Contribution Statement

Akihisa Tsuchimoto: Investigation (Equal), Writing – original draft (Equal)

Masashi Okubo: Conceptualization (Equal), Writing – review & editing (Equal)

Atsuo Yamada: Conceptualization (Equal), Writing – review & editing (Equal)

Conflict of Interest

The authors declare no conflict of interest in the manuscript.

Funding

Japan Society for the Promotion of Science: 20J23089

Data Creation and Utilisation Type Materials Research and Development Project: JPMXP1121467561

ANRI Fellowship

Footnotes

A. Tsuchimoto: ECSJ Student Member

M. Okubo: ECSJ Active Member

A. Yamada: ECSJ Fellow

References

• 1)  D. Larcher and J.-M. Tarascon, Nat. Chem., 7, 19 (2015).
• 2)  G. Crabtree, Science, 366, 422 (2019).
• 3)  H. Koga, L. Croguennec, M. Ménétrier, K. Douhil, S. Belin, L. Bourgeois, E. Suard, F. Weill, and C. Delmas, J. Electrochem. Soc., 160, A786 (2013).
• 4)  M. Sathiya, G. Rousse, K. Ramesha, C. P. Laisa, H. Vezin, M. T. Sougrati, M.-L. Doublet, D. Foix, D. Gonbeau, W. Walker, A. S. Prakash, M. Ben Hassine, L. Dupont, and J.-M. Tarascon, Nat. Mater., 12, 827 (2013).
• 5)  K. Luo, M. R. Roberts, R. Hao, N. Guerrini, D. M. Pickup, Y.-S. Liu, K. Edström, J. Guo, A. V. Chadwick, L. C. Duda, and P. G. Bruce, Nat. Chem., 8, 684 (2016).
• 6)  D.-H. Seo, J. Lee, A. Urban, R. Malik, S. Kang, and G. Ceder, Nat. Chem., 8, 692 (2016).
• 7)  T. Sudayama, K. Uehara, T. Mukai, D. Asakura, X.-M. Shi, A. Tsuchimoto, B. M. de Boisse, T. Shimada, E. Watanabe, Y. Harada, M. Nakayama, M. Okubo, and A. Yamada, Energy Environ. Sci., 13, 1492 (2020).
• 8)  N. Yabuuchi, M. Takeuchi, M. Nakayama, H. Shiiba, M. Ogawa, K. Nakayama, T. Ohta, D. Endo, T. Ozaki, T. Inamasu, K. Sato, and S. Komaba, Proc. Natl. Acad. Sci. U.S.A., 112, 7650 (2015).
• 9)  B. Mortemard de Boisse, G. Liu, J. Ma, S.-I. Nishimura, S.-C. Chung, H. Kiuchi, Y. Harada, J. Kikkawa, Y. Kobayashi, M. Okubo, and A. Yamada, Nat. Commun., 7, 11397 (2016).
• 10)  B. Mortemard de Boisse, S.-I. Nishimura, E. Watanabe, L. Lander, A. Tsuchimoto, J. Kikkawa, E. Kobayashi, D. Asakura, M. Okubo, and A. Yamada, Adv. Energy Mater., 8, 1800409 (2018).
• 11)  G. Assat and J.-M. Tarascon, Nat. Energy, 3, 373 (2018).
• 12)  N. Yabuuchi, Chem. Lett., 46, 412 (2017).
• 13)  M. Sathiya, A. M. Abakumov, D. Foix, G. Rousse, K. Ramesha, M. Saubanère, M. L. Doublet, H. Vezin, C. P. Laisa, A. S. Prakash, D. Gonbeau, G. VanTendeloo, and J.-M. Tarascon, Nat. Mater., 14, 230 (2015).
• 14)  A. Tsuchimoto, X.-M. Shi, K. Kawai, B. M. de Boisse, J. Kikkawa, D. Asakura, M. Okubo, and A. Yamada, Nat. Commun., 12, 631 (2021).
• 15)  K. Kawai, X.-M. Shi, N. Takenaka, J. Jang, B. M. de Boisse, A. Tsuchimoto, D. Asakura, J. Kikkawa, M. Nakayama, M. Okubo, and A. Yamada, Energy Environ. Sci., 15, 2591 (2022).
• 16)  P. M. Csernica, S. S. Kalirai, W. E. Gent, K. Lim, Y.-S. Yu, Y. Liu, S.-J. Ahn, E. Kaeli, X. Xu, K. H. Stone, A. F. Marshall, R. Sinclair, D. A. Shapiro, M. F. Toney, and W. C. Chueh, Nat. Energy, 6, 642 (2021).
• 17)  A. R. Armstrong, M. Holzapfel, P. Novák, C. S. Johnson, S.-H. Kang, M. M. Thackeray, and P. G. Bruce, J. Am. Chem. Soc., 128, 8694 (2006).
• 18)  W. Yin, A. Grimaud, G. Rousse, A. M. Abakumov, A. Senyshyn, L. Zhang, S. Trabesinger, A. Iadecola, D. Foix, D. Giaume, and J.-M. Tarascon, Nat. Commun., 11, 1252 (2020).
• 19)  M. Gu, I. Belharouak, J. Zheng, H. Wu, J. Xiao, A. Genc, K. Amine, S. Thevuthasan, D. R. Baer, J.-G. Zhang, N. D. Browning, J. Liu, and C. Wang, ACS Nano, 7, 760 (2013).
• 20)  H. Liu, K. J. Harris, M. Jiang, Y. Wu, G. R. Goward, and G. A. Botton, ACS Nano, 12, 2708 (2018).
• 21)  J. Hong, D.-H. Seo, S.-W. Kim, H. Gwon, S.-T. Oh, and K. Kang, J. Mater. Chem., 20, 10179 (2010).
• 22)  C. R. Fell, K. J. Carroll, M. Chi, and Y. S. Meng, J. Electrochem. Soc., 157, A1202 (2010).
• 23)  D. Eum, B. Kim, S. J. Kim, H. Park, J. Wu, S.-P. Cho, G. Yoon, M. H. Lee, S.-K. Jung, W. Yang, W. M. Seong, K. Ku, O. Tamwattana, S. K. Park, I. Hwang, and K. Kang, Nat. Mater., 19, 419 (2020).
• 24)  D. Kim, J. R. Croy, and M. M. Thackeray, Electrochem. Commun., 36, 103 (2013).
• 25)  R. A. House, U. Maitra, M. A. Pérez-Osorio, J. G. Lozano, L. Jin, J. W. Somerville, L. C. Duda, A. Nag, A. Walters, K.-J. Zhou, M. R. Roberts, and P. G. Bruce, Nature, 577, 502 (2020).
• 26)  A. Kilic, Ç. Odabaşı, R. Yildirim, and D. Eroglu, Chem. Eng. J., 390, 124117 (2020).
• 27)  P. M. Attia, A. Grover, N. Jin, K. A. Severson, T. M. Markov, Y.-H. Liao, M. H. Chen, B. Cheong, N. Perkins, Z. Yang, P. K. Herring, M. Aykol, S. J. Harris, R. D. Braatz, S. Ermon, and W. C. Chueh, Nature, 578, 397 (2020).
• 28)  Y. Liu, Q. Zhou, and G. Cui, Small Methods, 5, 2100442 (2021).
• 29)  C. Chen, Y. Zuo, W. Ye, X. Li, Z. Deng, and S. P. Ong, Adv. Energy Mater., 10, 1903242 (2020).
• 30)  A. D. Sendek, E. D. Cubuk, E. R. Antoniuk, G. Cheon, Y. Cui, and E. J. Reed, Chem. Mater., 31, 342 (2019).
• 31)  R. P. Joshi, J. Eickholt, L. Li, M. Fornari, V. Barone, and J. E. Peralta, ACS Appl. Mater. Interfaces, 11, 18494 (2019).
• 32)  N. Kireeva and V. S. Pervov, Batter. Supercaps. 3, 427 (2020).
• 33)  F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and É. Duchesnay, J. Mach. Learn. Res., 12, 2825 (2011).
• 34)  L. Breiman, Mach. Learn., 45, 5 (2001).
• 35)  M. Lundberg and S. Lee, Proceedings of the 31st International Conference on Neural Information Processing Systems, p.4768 (2017).
• 36)  G. Assat, D. Foix, C. Delacourt, A. Iadecola, R. Dedryvère, and J.-M. Tarascon, Nat. Commun., 8, 2219 (2017).
• 37)  J. R. Croy, K. G. Gallagher, M. Balasubramanian, Z. Chen, Y. Ren, D. Kim, S.-H. Kang, D. W. Dees, and M. M. Thackeray, J. Phys. Chem. C, 117, 6525 (2013).
• 38)  A. J. Perez, Q. Jacquet, D. Batuk, A. Iadecola, M. Saubanère, G. Rousse, D. Larcher, H. Vezin, M.-L. Doublet, and J.-M. Tarascon, Nat. Energy, 2, 954 (2017).
• 39)  Q. Jacquet, A. Perez, D. Batuk, G. Van Tendeloo, G. Rousse, and J.-M. Tarascon, Chem. Mater., 29, 5331 (2017).
• 40)  E. Boivin, N. Guerrini, R. A. House, J. G. Lozano, L. Jin, G. J. Rees, J. W. Somerville, C. Kuss, M. R. Roberts, and P. G. Bruce, Adv. Funct. Mater., 31, 2003660 (2021).
• 41)  D. Mohanty, J. Li, D. P. Abraham, A. Huq, E. A. Payzant, D. L. Wood, III, and C. Daniel, Chem. Mater., 26, 6272 (2014).