Predicting CO2 adsorption in KOH-activated biochar using advanced machine learning techniques

Accurately forecasting carbon dioxide (CO₂) adsorption in KOH-activated biochar is crucial for advancements in geoenergy engineering and environmental technology. This research aims to develop robust machine learning models to capture the intricate relationships influencing CO₂ adsorption, driven by variables like pressure, temperature, and the biochar's chemical and physical properties. We employed a comprehensive suite of machine learning methods, like convolutional neural networks, random forests, artificial neural networks, linear regression, ridge and lasso regressions, elastic net, support vector machines, decision trees, gradient boosting machines, k-nearest neighbors, light gradient boosting machines, extreme gradient boosting, CatBoost, and Gaussian process, to build predictive models. These models were trained and validated on a dataset of 329 data points, assessed through performance metrics and visualizations. The dataset's suitability was confirmed by Monte Carlo outlier detection. Detailed analysis, utilizing the Taylor Diagram and performance metrics, confirmed that SVR and CatBoost models achieved the highest accuracy in predicting CO₂ adsorption. Their superior performance is evidenced by high R² values of 0.9235 (SVR) and 0.9327 (CatBoost), coupled with low mean squared error values of 0.2207 (SVR) and 0.1942 (CatBoost). Sensitivity analysis further indicated all input parameters' correlation with CO₂ adsorption, while SHAP analysis identified pressure and temperature as critical factors. The results demonstrate the power of advanced machine learning methods, particularly CatBoost and SVR, in predicting CO₂ adsorption and offer valuable insights for industrial applications and future research efforts aimed at enhancing adsorption efficiency.