Lithium-rich manganese-based layered oxides (LLOs) are considered as next-generation promising cathode materials for lithium-ion batteries due to the high theoretical energy density (exceeding 1000 W·h·kg^-1) and low cost. However, their electrochemical performance is highly sensitive to the initial activation conditions, including initial charge cut off voltage, activation temperature, initial current density, and activation cycle number. The activation process involves the electrochemical transformation of Li₂MnO₃ domains into active phases, which is crucial for achieving high capacity but leads to irreversible oxygen release and structural degradation. Traditional optimization methods rely on single-factor experiments, which are inefficient and fail to address the complex coupling effects of multiple parameters. This study proposes a machine learning driven framework to optimize the activation conditions of LLOs, addressing the limitations of conventional approaches by enabling high-dimensional parameter search and predictive modeling. This work highlights the necessity of artificial intelligent optimization for enhancing the stability and performance of high-energy-density batteries.

The research constructed a dataset comprising nearly 100 samples from laboratory experiments and research articles, featuring key parameters such as Li₂MnO₃ domain ratio (x = 0.1-0.5), activation charge cutoff voltage (4.4-4.8 V), activation temperature (25-55 ℃), activation current density (20-220 mA·g^-1), and activation cycles (1-4). Data preprocessing included unit standardization and mean imputation for missing values. Six supervised learning regression models—Lasso Regression, Ridge Regression, Support Vector Regression (SVR), Decision Tree Regression (DTR), XGBoost, and CatBoost—were trained to predict the 50th-cycle capacity retention. Hyperparameter optimization was performed using grid search with 5-fold cross-validation, evaluated by metrics like R², MAE, MSE, and MAPE. The CatBoost model demonstrated superior prediction performance and was selected for further analysis. Feature importance was interpreted using SHapley Additive exPlanations (SHAP), and Bayesian optimization with Gaussian process regression was applied to identify optimal activation conditions by maximizing capacity retention. Experimental validation involved synthesizing LLOs with different Li₂MnO₃ ratios (e.g., 0.3Li₂MnO₃·0.7LiMn_0.42Ni_0.42Co_0.16O₂ and 0.5Li₂MnO₃·0.5LiMn_2.42Ni_0.42Co_2.16O₂), characterized by XRD and electrochemical performance testing.

The CatBoost model achieved the highest prediction accuracy (R² = 0.777, MSE = 3.00 × 10^-4), significantly outperforming linear models and other nonlinear algorithms. SHAP analysis revealed that Li₂MnO₃ domain ratio and activation charging cutoff voltage were the most influential features, with high voltages (>4.6 V) and high Li₂MnO₃ content negatively impacting capacity retention due to accelerated oxygen loss and structural degradation. Bayesian optimization identified distinct optimal conditions for different LLOs compositions: for a Li₂MnO₃ ratio of 0.3, the optimal parameters were 4.49 V charging cutoff voltage, 25 ℃ activation temperature, and 20 mA·g^-1 activation current density, yielding lowest capacity decay; for a ratio of 0.5, a higher temperature of 43.95 ℃ was required to fully activate the material while maintaining stability. Experimental validation confirmed the model's predictions, with capacity retention errors below 1% compared to experimental result. For instance, under optimized conditions, the 3/7 LLOs sample maintained 100.9% capacity retention after 50 cycles, versus 90.6% with conventional high-voltage activation. The results demonstrate that ML enables precise control of multi-parameter interactions, balancing capacity and stability by suppressing irreversible reactions.

This study successfully integrates machine learning with experimental research to optimize the electrochemical activation of lithium-rich cathodes. The CatBoost-based predictive model accurately captures nonlinear relationships between activation parameters and capacity retention, while Bayesian optimization efficiently identifies global optimal conditions in high-dimensional space. The work provides a data-driven strategy for designing high-performance LLOs, emphasizing the critical roles of voltage and temperature in structural evolution. This approach significantly reduces experimental costs and time, offering a generalizable framework for multi-parameter optimization in complex electrochemical systems. Future applications could extend to other battery materials and processes, advancing the development of intelligent materials design.