Probabilistic prediction of phosphate ion adsorption onto biochar materials using a large dataset and online deployment.

Phosphate (PO(III)) contamination in water bodies poses significant environmental challenges, necessitating efficient and accurate methods to predict and optimize its removal. The current study addresses this issue by predicting the adsorption capacity of PO(III) ions onto biochar-based materials using five probabilistic machine learning models: eXtreme Gradient Boosting LSS (XGBoostLSS), Natural Gradient Boosting, Bayesian Neural Networks (NN), Probabilistic NN, and Monte-Carlo Dropout NN. Utilizing a dataset of 2952 data points with 16 inputs, XGBoostLSS demonstrated the highest R (0.95) on new adsorbents. SHapely Additive exPlanations analysis showed that adsorption experimental conditions had the most significant impact (43%), followed by synthesis conditions (29%) and adsorbent characteristics (28%). Optimized conditions included an initial PO(III) concentration of 125 mg/L, carbon content of 11.5%, oxygen content of 23%, a contact time of 1440 min, a heating rate of 5 °C/min, 200 rpm, and a surface area of 410 m/g, using Ra-LDO adsorbent synthesized from rape cabbage feedstock. This study developed and presented a practical online framework for predicting PO(III) removal onto biochar using a web-based graphical user interface.