Deep learning has recently gained attention as a modeling tool in the environmental field, showing promising potential for predicting nonlinear interactions among various variables in complex systems. In water treatment processes in particular, contam...
Deep learning has recently gained attention as a modeling tool in the environmental field, showing promising potential for predicting nonlinear interactions among various variables in complex systems. In water treatment processes in particular, contaminant removal efficiency is influenced by intricate interdependencies among operational parameters. Conventional modeling approaches often fall short in accurately capturing these relationships. This dissertation aimed to address these limitations by integrating deep learning models with experimental adsorption studies to improve the predictive accuracy of contaminant removal from aqueous solutions.
In this study, a range of functionalized adsorbents were synthesized to remove specific pollutants, including nitrate, diclofenac, naproxen, and Cu(II) ions. The adsorbents included quaternized poly(amidoamine) (PAMAM) dendrimers, Fe(III)- or Ca(II)-doped β-cyclodextrin/chitosan cryogels, and amine-functionalized cellulose beads. Each material was designed to chemically or physically interact with its target contaminant. To evaluate their removal performance, batch adsorption experiments were conducted under various conditions, and column studies were also performed to investigate dynamic adsorption behavior under flow-through conditions.
To ensure experimental rigor and broad parameter coverage, statistical experimental design methods such as Central Composite Design (CCD) and Box-Behnken Design (BBD) were applied. These approaches enabled the generation of structured datasets suitable for machine learning applications.
Based on these datasets, various data-driven modeling techniques were developed and evaluated. Multilayer perceptron (MLP) neural networks were applied to model the influence of variables such as concentration, pH, temperature, and adsorbent dosage on removal efficiency. Radial basis function networks (RBFNs) were employed to capture localized nonlinear patterns. Support vector regression (SVR), a classical non-neural machine learning approach, was used as a benchmark model for comparative evaluation. For modeling breakthrough curves obtained from column experiments, bidirectional long short-term memory (BiLSTM) networks were implemented to capture time-dependent adsorption dynamics.
This research was conducted through a consistent framework that combined functional adsorbent synthesis, systematic experimental design, and deep learning-based modeling. The results demonstrated that deep learning models, particularly neural networks, could effectively replace traditional polynomial-based regression methods while offering improved generalization performance. In addition, the models enabled variable importance assessment and process optimization, contributing to enhanced predictive capacity across diverse operating conditions and supporting the identification of optimal treatment parameters.
By integrating AI-based modeling with materials science and environmental engineering, this dissertation has shown the practical feasibility of applying deep learning to adsorption-based water treatment. Moreover, the study demonstrated that well-trained deep learning models could serve as efficient surrogates for large-scale experimentation, potentially reducing the experimental burden. The methodology presented here may be extended to other environmental applications and could serve as a foundation for future efforts in material design, process optimization, and real-time operational control. Ultimately, this work aims to support the advancement of more efficient and adaptable water treatment systems.