Advanced hybrid evaluation of water treatment using porous materials for adsorption separation via machine learning and mechanistic models

This study presents an advanced hybrid evaluation approach for predicting chemical concentration (C) in adsorption-based water treatment processes using a combination of tree-based machine learning models and Massively Parallel Hyperparameter Tuning. The objective is to accurately model the nonlinear relationships between spatial input parameters (x and y) and concentration outputs within a complex porous-material system. Three ensemble learning algorithms—Random Forest (RF), Gradient Boosting (GB), and Extra Trees (ET)—were systematically optimized and assessed to determine their suitability for high-precision concentration prediction. The tuning framework enabled extensive exploration of hyperparameter space, significantly enhancing model performance. Among the tested models, Extra Trees (ET) demonstrated outstanding predictive capability, achieving an R 2 value of 0.99924, along with the lowest MAPE (2.52675E-02) and MAE (5.59418E-03). These metrics confirm the ET model’s exceptional ability to capture subtle nonlinear trends and complex interactions inherent to adsorption-driven systems. In comparison, RF and GB also achieved strong performance but fell short of ET in both accuracy and robustness for the data analysis. The results highlight the effectiveness of parallelized hyperparameter optimization in improving predictive reliability for chemically intricate processes. This work underscores the value of machine-learning-assisted modeling for water treatment applications and provides a scalable framework that can support process design, operational decision-making, and further mechanistic integration in environmental and chemical engineering.