Machine Learning Improves Prediction of Emerging Inorganic Contaminant Adsorption Removal Compared With Simple and Multiple Linear Regression

Accurate prediction of emerging inorganic contaminant adsorption is important for accelerating the screening of adsorbent materials for water and wastewater treatment. This study directly compared simple linear regression (SLR), multiple linear regression (MLR), random forest, and extreme gradient boosting (XGBoost) using the same adsorption dataset and common evaluation metrics. The final dataset comprised (n=302) observations, partitioned into training (75%), validation (15%), and test (15%) subsets after data cleaning and normalization. SLR and MLR showed poor predictive performance (R² = −0.08 and −0.30, respectively), whereas the nonlinear ensemble models substantially reduced prediction error. On the test subset, random forest achieved R² = 0.51, MAE = 2.44, and RMSE = 3.10, while XGBoost achieved R² = 0.78, MAE = 2.07, and RMSE = 3.90. Thus, XGBoost provided the highest explained variance and the lowest mean absolute error, whereas random forest produced the lowest test RMSE. Feature-importance analysis identified nitrogen content and pore volume as the most influential predictors, followed by oxygen content, hydrogen content, ash content, carbon content, specific surface area, and pH. The results demonstrate that nonlinear tree-based models provide a more informative predictive framework than conventional linear regression for this dataset, although external validation and inclusion of operational variables such as initial contaminant concentration and adsorbent dosage are needed before broad practical application.