Data-augmented machine learning improves water treatment design: Precise prediction of PPCPs reaction with reactive...

Radical-mediated advanced oxidation and/or reduction processes (AOPs/ARPs) have shown remarkable efficacy in degrading organic pollutants in wastewater, accurate prediction of radical-pollutant reaction kinetic and thorough mechanistic understanding are critical for optimizing these processes. Although machine learning (ML) has emerged as a powerful predictive tool, its performance is often limited by small datasets, leading to overfitting and poor mechanistic interpretation. To overcome these limitations, we develop a novel data augmentation framework based on variational autoencoder (VAE) strategy, using carbon dioxide radical (CO₂^•-) based reductive process as a model system. Our results demonstrate that VAE-generated synthetic data significantly enhances model performance, particularly for nonlinear models, highly improving test-set prediction accuracy (R²) by 0.15-0.29. The optimized VAE-Artificial Neural Network (VAE-ANN) model achieves exceptional predictive capability with R² of 0.99 and 0.88 on the training and test data, respectively. Mechanistic analysis identifies HOMO-LUMO gaps (E_GAP), molecular hardness (S), and surface area of molecular electrostatic potential (ESP_{pos per}) as the most critical molecular factors governing the reaction kinetics. Notably, VAE-ML framework also shows excellent generalizability, with excellent performance when extended to hydroxyl radical based AOPs. The present study not only provides a generalizable data augmentation strategy to address small-sample limitations in ML, but also offers molecular-level insights that enable optimization of water treatment applications.