Researchers from King Khalid University and partner institutions apply AI-based regression models—including Multilayer Perceptron (MLP), Gaussian Process Regression (GPR), and Polynomial Regression (PR)—to computational fluid dynamics (CFD) datasets of adsorption processes. After preprocessing with a local outlier factor and gradient-based hyperparameter tuning, the MLP achieves superior predictive performance (R2=0.999, RMSE=0.583), demonstrating strong potential for environmental process optimization.
Key points
MLP regression on CFD-derived adsorption data achieves R2=0.999 and RMSE=0.583, outperforming GPR and PR.
Preprocessing uses Local Outlier Factor for data cleaning and Min-Max scaling for normalization.
Gradient-based hyperparameter optimization and five-fold cross-validation validate MLP’s robustness (AARD%=2.56%).
Why it matters:
This approach provides rapid, high-accuracy solute concentration predictions, enhancing adsorption-based water purification and resource-efficient environmental monitoring.
Q&A
What is adsorption in water treatment?
How does computational fluid dynamics (CFD) generate training data?
Why use Local Outlier Factor (LOF) for outlier detection?
What is gradient-based hyperparameter optimization?
Read full article
Academy
Adsorption in Porous Materials
Adsorption is a surface phenomenon where molecules from a fluid (liquid or gas) adhere to the surface of a solid material, often referred to as the adsorbent. Porous materials—such as activated carbon, zeolites, and mesoporous silica—possess a high internal surface area that enhances their ability to capture and retain solute molecules. In water treatment, adsorption is widely used to remove organic pollutants, heavy metals, and dyes from wastewater streams.
The adsorption process relies on intermolecular forces, including van der Waals interactions, hydrogen bonding, and electrostatic attractions between the adsorbent surface and the target solute. Key parameters affecting adsorption performance include:
- Surface area: Larger area provides more binding sites.
- Pore size distribution: Determines accessibility for different molecule sizes.
- Chemical functionality: Surface modifications introduce functional groups that improve affinity for specific solutes.
- Operational conditions: Temperature, pH, and flow rate influence adsorption kinetics and equilibrium.
Advanced modeling techniques, such as molecular simulations and computational fluid dynamics (CFD), help researchers understand how solute molecules distribute within porous media under various conditions. These simulations can reveal concentration gradients, diffusion pathways, and adsorption isotherms essential for designing efficient treatment systems.
Multilayer Perceptron (MLP) in Regression Tasks
A Multilayer Perceptron (MLP) is a class of feedforward artificial neural network composed of an input layer, one or more hidden layers, and an output layer. Each layer contains nodes (neurons) that apply a weighted sum of inputs followed by a nonlinear activation function.
For regression tasks—such as predicting solute concentration—MLPs learn a mapping from input features (e.g., spatial coordinates x and y) to a continuous output (solute concentration). The process involves:
- Data normalization: Input features are rescaled, often using Min-Max scaling, to ensure efficient training.
- Training: The network adjusts weights and biases by minimizing a loss function (e.g., mean squared error) through backpropagation and an optimization algorithm like Adam or stochastic gradient descent.
- Validation: Performance is evaluated on unseen data via metrics like the coefficient of determination (R²) and root mean square error (RMSE).
- Hyperparameter tuning: Gradient-based methods or grid search optimize parameters such as learning rate, number of layers, and neuron count.
MLPs are powerful tools for capturing complex, nonlinear relationships in multiscale datasets, making them well-suited for modeling physical processes like adsorption. When combined with robust preprocessing and validation techniques, they can deliver high-accuracy predictions essential for process optimization and environmental engineering.
