# Uncertainty‐Quantified Primary Particle Size Prediction in Li‐Rich NCM Materials via Machine Learning and Chemistry‐Aware Imputation

## TL;DR

This study uses machine learning to predict the particle size of Li-rich NCM battery materials, even with incomplete data, showing that sintering conditions are key factors.

## Contribution

A novel ML framework with chemistry-aware imputation and uncertainty quantification for predicting particle size in Li-rich NCM materials.

## Key findings

- MatImpute-based NGBoost model achieved a test R2 of 0.866 and calibration error of 0.133.
- Sintering temperature and time were identified as dominant factors in particle size prediction.
- Experimental validation showed predictions within 0.13 µm of measurements with uncertainties near 1.5.

## Abstract

Lithium‐rich Nickel–Cobalt–Manganese (Li‐rich NCM) materials are promising cathodes for lithium‐ion batteries, where electrochemical performance is sensitive to primary particle size. Yet, efforts to apply machine learning (ML) for particle size prediction are constrained by incomplete literature data. This study applies imputation methods—MatImpute, K‐nearest neighbors, multivariate imputation by chained equations, and Mean—to complete the datasets, followed by training a Natural Gradient Boosting (NGBoost) model to predict primary particle size with quantified uncertainty. Two training strategies are evaluated: one including entries with imputed target values and another excluding them. Both strategies are tested on a fully observed dataset. The MatImpute‐based NGBoost model achieves the highest accuracy, with a test R2 of 0.866 and a calibration error of 0.133. Feature analysis identifies second sintering temperature and first sintering time as dominant factors, with composition showing minimal influence, consistent with prior experimental reports that sintering parameters drive sub‐micron grain growth through atomic mobility and grain coarsening. Experimental validation shows most predictions within 0.13 µm of measurements and normalized uncertainties near 1.5. These findings demonstrate that robust imputation and uncertainty quantification enhance ML‐based particle size prediction and confirm that sintering conditions, rather than stoichiometry, govern microstructural evolution in Li‐rich NCM materials.

This study demonstrates a machine learning framework that predicts the primary particle size of Lithium‐rich Nickel‐Cobalt‐Manganese (Li‐rich NCM) materials from synthesis conditions, even with incomplete literature data. By combining chemistry‐aware imputation with uncertainty‐quantified modeling, it identifies sintering temperature and time as dominant factors, enabling more reliable, data‐driven optimization of microstructure for high‐performance lithium‐ion batteries.