Comparative Study of Machine Learning Methods for Modeling Graphene-Based Adsorption in Water Treatment

Area of Science:

• Environmental Science
• Materials Science
• Data Science

Background:

• Polynomial regression (PR) is widely used for predicting pollutant removal efficiency but lacks flexibility.
• Machine learning methods like support vector machines (SVM) and artificial deep learning neural networks (ANNDL) show promise but require optimization for small datasets.
• Adsorption studies for water pollutant treatment often involve small datasets, necessitating robust modeling approaches.

Purpose of the Study:

• To evaluate the performance of Polynomial Regression (PR), Support Vector Machines (SVM), and Artificial Deep Learning Neural Networks (ANNDL) for predicting water pollutant removal efficiency using small datasets.
• To compare the flexibility and tolerance of these methods concerning data size and parameter limitations.
• To provide recommendations for optimizing adsorption systems and experimental design in water treatment.

Main Methods:

• Utilized four published datasets (graphene-related nanomaterials) with small data sizes (20-30 samples).
• Applied Polynomial Regression (PR), Support Vector Machines (SVM) with a radial basis function kernel, and Artificial Deep Learning Neural Networks (ANNDL) using the TensorFlow framework.
• Evaluated model performance and reliability under cross-validation, focusing on data size and parameter sensitivity.

Main Results:

• Polynomial Regression (PR) showed reliability but was highly sensitive to data size and parameter limitations.
• Support Vector Machines (SVM) with a radial basis function kernel and Artificial Deep Learning Neural Networks (ANNDL) demonstrated superior performance and greater tolerance to variations in data size and parameters compared to PR.
• Both SVM and ANNDL proved more robust for predicting pollutant removal efficiency with limited data.

Conclusions:

• SVM and ANNDL are more suitable than PR for modeling pollutant removal efficiency in small-sized datasets common in adsorption studies.
• The choice of modeling technique significantly impacts the reliability and optimization of water treatment systems.
• Future research should focus on leveraging advanced machine learning techniques for robust experimental design and effective water pollutant treatment.