Characterization and Prediction of Coal Particle Size Distribution before and after Cleaning: An Integrated Experimental and Machine Learning Approach

The particle size distribution (PSD) of coal during postmining activities, such as transportation and cleaning processes, exhibits a crucial correlation with methane emission characteristics. However, systematic investigations addressing the combined effects of diverse geological conditions and processing techniques remain insufficient. This study conducted comprehensive experimental analyses on coal samples from 16 representative mines in 9 China’s major coal bases, integrating mechanical sieving, dynamic image analysis, and morphological characterization to quantify full-scale PSD and shape parameter variations before and after cleaning. And machine learning models were developed to establish predictive relationships between coal properties, process parameters, and resulting granularity characteristics. The results show that fine coal particles (<1 mm) constitute a low proportion in raw coal, while medium-sized particles (1–30 mm) dominate cleaned products due to their optimal processing characteristics. The cleaning processes transform initially unimodal left-skewed PSD into distinct bimodal patterns. This shift is characterized by a significant increase in the proportion of fine particles (<1 mm) and the angularity of large particles (>10 mm), primarily due to selective removal and morphological modification mechanisms. Coal particle shape analysis reveals strong size-dependent morphological characteristics, with circularity showing concentrated distribution trends in coarse particles, while the ellipse ratio maintains dispersed patterns across all size ranges. And the cleaning process promotes gradual shape optimization from underground raw coal to clean coal. Low-rank bituminous coal maintains larger average particle sizes across all size fractions compared to higher-rank coals. Lower firmness coefficients (f ≤ 0.3) correlate with a higher proportion of fine particles and smaller average sizes, and combined cleaning techniques enable more precise size distribution control than single-technique processes. For predictive modeling, ensemble methods for Random Forest and Gradient Boosting Decision Tree are particularly suitable for predicting precleaning characteristic particle sizes, achieving a high accuracy of R2 ≥ 0.972. And the XGBoost algorithm demonstrates superior performance for accurately estimating size parameters after cleaning. Feature importance analysis confirms the inherent coal particle size as the most influential factor, contributing over 70% to the model’s output. This research could provide theoretical references for predicting coal PSD under varying conditions to optimize cleaning operations and advance methane emission quantification through precise particle size prediction.