Screening of Next Generation Ni-Rich Cathode Materials with High Electrochemical Properties via Machine Learning

Ni-rich layered cathode materials are promising candidates to satisfy high energy density and high voltage requirements, but they suffer from degradation during cycling. Ni-rich layered cathodes, such as LiNi_xCo_yMn_zO₂ (NCM) and LiNi_xCo_yAl_zO₂ (NCA), were extensively investigated to minimize the proportion of Co, which caused price volatility and supply chain problems. Recently, LiNi_xCo_yMn_zAl_wO₂ (NCMA), with the minimum Co content and maximum Ni content, was studied owing to its advantages of low manufacturing cost and high theoretical capacity and working potential. However, as the Ni content increases in the cathode, its commercialization is challenging owing to (1) performance degradation and (2) safety issues during electrochemical cycling. Therefore, the development of next-generation Ni-rich cathodes is necessary to resolve these problems.<br/>Density functional theory (DFT) calculations are widely implemented in designing electrode materials owing to their remarkable accuracies and reliabilities. However, DFT calculations require significant computational resources to calculate all possible combinations of candidate materials. This limits the scope of the search for novel materials and the possibilities of fundamental research and development. Thus, to efficiently design next-generation battery materials, narrowing the large design space for cathode materials using artificial intelligence (AI) is necessary. And although various battery-material studies aim to yield cathodes with high voltage and small volume changes, few studies focus on developing prediction models of the average voltage and volume changes of novel materials.<br/>In this study, we developed a machine-learning-based surrogate model to predict the average voltage and volume changes of Ni-rich cathodes with various dopants (LiNi_0.85D^’_xD^’’_{(0.15 – x)}O₂) to determine ideal cathode materials with excellent electrochemical properties. To construct the training database, data regarding 4,401 materials were obtained from the Materials Project. Thirty-three elements were implemented as candidate dopants, suggesting 1,617 potential cathode materials.<br/>We improved the accuracy of predictive models through feature engineering including feature addition and data pre-processing. As a result of adjusting the hyperparameters with Bayesian optimization, the performance improved significantly from the initial to the final model (average voltage: R² = 0.801 to 0.873, MAE = 0.404 to 0.323 V; volume change: R² = 0.241 to 0.561, MAE = 41.030% to 2.890%).<br/>Using the constructed model, we identified 107 candidate materials. The following are the criteria for selecting cathode materials: (1) The gravimetric energy density should be 〉875 mWh/g based on the capacity of Ni-rich compounds (200-250 mAh/g). (2) Based on the potentials (2.7–4.3 V) of recently proposed Ni-rich layered materials, the average voltage should be 〉3.5 V. (3) The volume change should be 〈7% based on the Ni-rich compounds (2–9%). 969 materials satisfy criterion (1), 541 satisfy criteria (1) and (2), and 107 satisfy criteria (1)–(3).<br/>The constructed platform may be employed to determine ideal Ni-rich cathode materials with different elemental ratios and compositions, with significantly reduced computational and experimental costs. The model was validated using DFT calculations. We identified 101 Co-free compounds among the candidates and presented a strategy for material selection that could overcome resource limitations. The top ten materials displayed high potentials and capacities and low volume changes without Co, suggesting the potential of developing Co-free cathode materials. The platform used in this study accelerated the screening of unknown materials without extensive calculations. These next-generation cathode candidates may be used in future experimental research regarding cathode material development or predictive models based on constructed databases.