Machine Learning-Based Prediction of Coal Particle Size Effects on CO<sub>2</sub> Adsorption.

Abstract

As global carbon emissions continue to rise, intensifying climate change and the greenhouse effect, achieving carbon peaking and carbon neutrality has become a pivotal goal in global climate governance. Carbon capture, utilization, and storage (CCUS) has become a crucial technology for achieving these goals and improving energy recovery, making it one of the key pathways to carbon neutrality. This study investigates the CO₂ adsorption characteristics of three different coal samples across various particle sizes, as determined through CO₂ isothermal adsorption experiments. We employed four machine learning models─XGBoost, SVM, LSTM, and CNN─trained and validated using two data preprocessing methods: sequential sorting and random sorting. A CO₂ adsorption capacity prediction model was established, with coal particle size as the input variable. The findings indicate that the model trained with randomly sorted data demonstrates significantly better prediction accuracy on the test set compared to the model trained with sequentially sorted data, with an average <i>R</i>² improvement of approximately 0.1. This indicates that randomizing the data effectively eliminates potential dependencies on time or particle size sequences, facilitating the model to grasp broader adsorption patterns and evade overfitting. Additionally, the absolute and square error indices show marked differences under different ranking methods for the same model, emphasizing the importance of selecting appropriate models based on specific circumstances. Through analyses using Taylor diagrams and the TOPSIS method, it was found that the random ranking model outperforms the sequential ranking model. The SVM model performs best in the Taylor diagram analysis, while the CNN model achieves the highest comprehensive evaluation in the TOPSIS method. SHAP value analysis reveals that the adsorption capacity for CO₂ in coal samples sized between 60 and 80 mesh is the most globally important factor for predicting the adsorption capacity of CO₂ in coal samples with a particle size exceeding 200 mesh. This finding highlights that coal's pore structure and adsorption kinetics are crucial factors influencing its CO₂ adsorption capacity. Overall, the machine learning model effectively predicts the CO₂ adsorption amount of coal, simulates the actual changes in the adsorption process, uncovers the CO₂ adsorption mechanism and critical influencing factors of coal, and enhances resource utilization efficiency.