Coupling generative and predictive machine learning algorithms to enhance haloacetonitriles prediction in small water...

Area of Science:

• Environmental Science
• Water Quality
• Computational Chemistry

Background:

• Small water systems face data scarcity challenges in modeling emerging disinfection byproducts (EDBPs).
• Existing modeling approaches are often limited by insufficient data, hindering effective EDBP surveillance.
• Resource-limited systems require cost-effective solutions for accurate water quality monitoring.

Purpose of the Study:

• To introduce a novel framework coupling generative algorithms (GAs) and machine learning algorithms (MLAs) to overcome data scarcity in EDBP modeling.
• To evaluate the performance of various GA-MLA combinations for EDBP classification and prediction.
• To assess the cost-effectiveness and viability of the proposed framework for small water systems.

Main Methods:

• Evaluated three generative algorithms: Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP), Variational Autoencoder (VAE), and Add Noise.
• Integrated GAs with four machine learning algorithms: Generalized Regression Neural Network (GRNN), eXtreme Gradient Boosting (XGBoost), Gaussian Process Regression (GPR), and Support Vector Regression (SVR).
• Utilized full-scale sampling data for model training and validation, performing sensitivity analysis on generated data volume.

Main Results:

• Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) generated data with the highest consistency with original samples.
• Coupled GA-MLA models significantly improved classification accuracy, precision, and specificity for dibromoacetonitrile exceedance compared to baseline MLAs.
• The optimal coupling (CWGAN-GP + GRNN) increased classification accuracy from 74.1% to 81.5% and specificity from 45.8% to 62.5%.
• CWGAN-GP and VAE significantly enhanced prediction accuracy for dichloroacetonitrile concentration across different MLAs, reducing Mean Squared Error (MSE).
• The GA-MLA framework reduced modeling expenses by up to 49.7%.

Conclusions:

• The proposed GA-MLA framework effectively mitigates data limitations in EDBP modeling for small water systems.
• This coupled approach provides a viable and economical pathway for EDBP surveillance, reducing reliance on extensive sampling.
• The study demonstrates the potential of generative AI in enhancing water quality monitoring and management in resource-limited settings.