Machine-Learning-Accelerated Development of High-Nickel NCM Cathodes via Multivariable Co-optimization

This study presents a machine-learning-based active-learning framework for optimizing high-nickel NCM cathode materials using a large-scale industrial dataset. Drawing from 3,019 pilot-scale experiments accumulated over two years, we utilized 706 high-quality samples for model development, capturing rich process variability under real manufacturing conditions. The framework was tested on a commercially important high-nickel NCM (LiNixCoyMn1-x-yO2, x ≥ 80%) cathode material containing 94% Ni, for which only a severely limited dataset of 18 samples was available. Using a Gradient Boosting model and iterative active learning, we achieved a discharge capacity of 228.3 mAh/g with only 38 experiments─reducing experimental effort by 94% compared to traditional methods. The model successfully leveraged human design biases to guide exploration beyond expert heuristics, discovering nonintuitive yet effective process conditions. By harnessing large, historically fragmented datasets, this work demonstrates a scalable approach for accelerating battery materials optimization in industrial environments.