Explainable metaheuristic-optimized XGBoost models for accurate prediction of CO2 capture by graphene-based adsorbents

The continued rise in atmospheric CO 2 levels is a leading driver of global climate change, necessitating the development of effective carbon capture strategies. Among these, adsorption-based CO 2 removal using engineered nanomaterials offers a promising approach due to its energy efficiency, modularity, and tunable surface chemistry. In this study, we present a hybrid machine learning (ML) framework that combines the predictive accuracy of eXtreme Gradient Boosting (XGBoost) with the global optimization strength of Differential Evolution (DE) to model CO 2 uptake by graphene-based adsorbents. The dataset, derived from controlled thermogravimetric measurements of CO 2 adsorption onto GO, rGO, and amine-functionalized GO variants, was used to train and evaluate multiple ML models. Following data pre-processing, including feature encoding and normalization, the DE-XGBoost model was benchmarked against several baseline and advanced ML algorithms, including linear regression, multilayer perceptron (MLP), random forest (RF), and non-optimized XGBoost. The DE-XGBoost model outperformed all other approaches, achieving R 2 , RMSE, and MAE values of 0.986,1.9 mg/g, and 1.2 mg/g, respectively. SHAP (SHapley Additive exPlanations) analysis consistently indicates that adsorption time is the most influential feature across different model settings and scenarios. This study demonstrates the power of metaheuristic optimization in enhancing both the accuracy and interpretability of ML models, providing a promising data-driven framework for accelerating CO 2 adsorbent performance prediction. • A novel PGB-LTH nanocomposite for the adsorptive removal of dyes was prepared • Ensemble machine learning models using a Decision Trees algorithm to predict dyes adsorption using the synthesized nanocomposite were developed. • Thorough comparison between the developed models and other machine learning models is provided.