Machine learning vs. statistical model for prediction modeling and experimental validation: Application in...

Area of Science:

• Environmental Engineering
• Water Resource Management
• Geochemistry

Background:

• Permeable reactive barriers (PRBs) are effective in-situ groundwater remediation technologies.
• PRB design hinges on critical factors like barrier width and reactive material selection.
• Arsenic contamination in groundwater poses significant environmental and health risks.

Purpose of the Study:

• To investigate beaded coal mine drainage sludge (BCMDS) as a reactive material for arsenic adsorption in PRBs.
• To compare traditional design methods with machine learning (ML) approaches for determining PRB width.
• To optimize ML models for accurate prediction of the mass transfer zone width (W_MTZ).

Main Methods:

• Utilized traditional column experiments and empirical formulas for PRB width determination.
• Employed machine learning (ML), specifically the XGBoost algorithm, for predicting W_MTZ using data from existing literature.
• Validated ML predictions against experimentally derived W_MTZ values and compared with multiple linear regression (MLR).

Main Results:

• The XGBoost ML model achieved high accuracy in predicting W_MTZ (R² = 0.97, RMSE = 0.15).
• ML predictions showed a low error rate of 7.04% when validated against experimental data.
• Multiple linear regression (MLR) exhibited a significantly higher error rate of 39.43%.

Conclusions:

• Machine learning offers superior accuracy and efficiency for PRB design compared to traditional methods, especially for complex contaminant scenarios.
• ML effectively accounts for material, pollutant, and environmental factors in predicting barrier width.
• Further development of ML models holds promise for broader application in groundwater remediation design.