Citation

Li Xiangyue, Zhu Dexin, Pan Kunmin, Zhou Xiaoye, Zhu Jiaming, Wang Yingxue, Ren Yongpeng, Wu Hong-Hui. Identifying key determinants of discharge capacity in ternary cathode materials of lithium-ion batteries[J]. Chin. Chem. Lett., 2025, 36(5): 109870  

Xiangyue Li^a,1, Dexin Zhu^b,1, Kunmin Pan^a,*, Xiaoye Zhou^c, Jiaming Zhu^d,e, Yingxue Wang^f,*, Yongpeng Ren^a, Hong-Hui Wu^b,d,*     

Lithium-ion batteries (LIBs) dominate the energy storage market due to their high energy density, long cycle life, and environmental compatibility [1-6]. Among cathode materials, Li-Ni-Co-Mn (LNCM) ternary materials, comprising Ni, Co, and Mn, are distinguished for their high energy density, durability, and environmental sustainability [7-10]. The initial discharge capacity (IC) of LNCM materials is a critical performance metric, affecting cycle life and indicative of overall performance [11,12]. Moreover, the discharge capacity at the 50^th cycle (EC) stands out as a critical parameter indicative of a battery's cycle endurance. Hence, the pursuit and discovery of strategies to enhance the IC and EC performances have become prominent subjects of inquiry in the realm of battery.

Discharge capacity, a pivotal metric for lithium-ion ternary battery performance, mirrors the available electric charge [13,14]. Doping is a prevalent approach to augment discharge capacity by optimizing LNCM material properties [15-18], and optimizing crystal-electronic structures to enhance capacity and cycle stability [19-21]. It increases specific surface area, boosting active material utilization and, consequently, battery discharge capacity [21,22]. Doping also optimizes the electronic structure, quickening lithium-ion diffusion, decreasing internal resistance, and uplifting overall battery efficacy [19]. The recent synthesis of AI and materials science has sparked keen interest due to their synergistic significance [23,24]. Machine learning, a key component of AI, offers pathways for optimizing and developing material properties [25,26]. Neural networks, for example, fulfill the need for high-efficiency solid-state electrolytes (SSE) materials [27] and accurately forecast polymer electrolyte ion conductivity [28]. Similarly, machine learning can predict the discharge capacity of LNCM ternary materials, guiding improvements in LIBs' performance.

This study aims to identify key material descriptors influencing IC and EC of LNCM ternary materials using machine learning. Data on IC and EC for LNCM ternary materials, doped with various single elements, were collected from academic publications. Subsequently, 229 material descriptors were constructed using domain knowledge and XenonPy [29]. An effective feature screening method was developed, integrating the Pearson correlation coefficient (PCC), importance coefficient, optimal subset method [30], and SHapley Additive exPlanations (SHAP) plots. This method pinpointed key features with high prediction accuracy for both IC and EC of LNCM materials [31], the workflow was shown in Fig. 1. SHAP plots elucidated potential relationships between these features and the target performance.

Figure 1

	Download: JPG larger image
Fig. 1. A workflow of screening the key factors in the discharge capacity of LNCM.

This work integrated dual feature sets originating from XenonPy calculations and domain-specific knowledge, summarized in Tables S1 and S2 (Supporting information). This study compiled a dataset of 237 single-element doped LNCM materials, as shown in Tables S3 and S4 (Supporting information), and normalized the feature set by Eq. S8 (Supporting information) [32-34]. After screening for feature independence, importance, and interpretation (Figs. S1 and S2 in Supporting information), we distilled down to 13 vital features. To derive the optimal subset, the optimal subset method is adopted for screening. As shown in Fig. 2a, observing a declining trend in PCC beyond four features, we concluded that the optimal subset comprises four attributes: var: ground_state_magnetic_moment (VMM), MTE, Vmax, and CD.

Figure 2

	Download: JPG larger image
Fig. 2. Selecting the optimal feature set and regression model. (a) PCC value was obtained by the optimal subset selection method. (b) The performance of 5 regression models with PCC, MAE, and RMSE values.

According to the "no free lunch" law, it is impractical to use a single algorithmic model to solve all machine learning problems [30,35-37]. Therefore, this study selects 5 typical models by ten-fold cross-validation [38-40], as shown in Fig. 2b. A series of indexes, PCC, MAE, and RMSE, were employed to gauge model performance [41-44]. Inspection of the results in Fig. 2b demonstrates that the XGB model consistently exhibits optimal performance across all three indexes, with PCC at 0.9034 and MAE and RMSE measured at 13.8033 and 18.9776 mAh/g, respectively. This indicates that through its exploration of the underlying relationship between the optimal set of features and the target performance, the XGB model achieves accurate and low-bias predictions. Consequently, the XGB model is selected as the preferred output model for predicting IC.

After feature engineering screening, four features demonstrating high predictive accuracy for target performance were identified, and Fig. 3a visually illustrates their predictive strength. By inputting the optimal subset of features into various regression models to select the model demonstrating the best performance, as illustrated in Fig. 2a, the XGB model was chosen as the optimal output model for predicting IC. Similarly, following the same rationale, the RF model was selected as the top-performing model for forecasting EC (Fig. 3b). After 100 random tests, the target performance yielded PCC = 0.9012, MAE = 13.9376 mAh/g, and RMSE = 19.1462 mAh/g. Beyond the notable predictive capability for IC, this screening approach also demonstrates heightened performance in predicting EC. Following 100 random tests, the performance indexes were PCC = 0.8461, MAE = 15.2549 mAh/g, and RMSE = 20.8239 mAh/g, highlighting the robustness of the model.

Figure 3

	Download: JPG larger image
Fig. 3. The scatterplot of the IC optimal model and EC optimal model: The scatterplot of (a) IC model and (b) EC model based on the same feature set.

The study further analyzed the impact of specific features on target performance using SHAP plots. Figs. 4a-d presented SHAP scatter plots, where Figs. 4a and c highlighted the positive correlation between Vmax and MTE with target performance, suggesting improved performance with higher Vmax and MTE values. Conversely, Figs. 4b and d associate CD and VMM with target performance, revealing a negative relationship as well.

Figure 4

	Download: JPG larger image
Fig. 4. The relationship between each feature and the IC based on the SHAP value diagram: The scatterplot of (a) Vmax, (b) CD, (c) MTE, and (d) VMM with their SHAP value, respectively.

In summary, this research successfully identified four key features significantly affecting the IC and EC of LNCM ternary materials. By examining the potential relationships between these features and the target performance using SHAP, we contribute to advancing battery technology and energy storage solutions. Furthermore, this methodology offers a basis for exploring performance prediction and optimization strategies in other battery materials.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Xiangyue Li: Resources, Software, Writing – original draft, Writing – review & editing. Dexin Zhu: Software, Supervision, Writing – review & editing. Kunmin Pan: Funding acquisition, Resources, Supervision, Writing – review & editing. Xiaoye Zhou: Investigation, Supervision. Jiaming Zhu: Investigation, Validation. Yingxue Wang: Data curation, Software, Funding acquisition. Yongpeng Ren: Methodology, Visualization. Hong-Hui Wu: Funding acquisition, Resources, Supervision, Writing – review & editing.

This study was financially supported by the National Natural Science Foundation of China (Nos. 52122408, 52071023), the Program for Science & Technology Innovation Talents in the University of Henan Province (No. 22HASTIT1006), the Program for Central Plains Talents (No. ZYYCYU202012172), the Ministry of Education, Singapore (No. RG70/20) and the Opening Project of National Joint Engineering Research Center for Abrasion Control and Molding of Metal Materials, Henan University of Science and Technology (No. HKDNM201906).