Variational Autoencoder-Regulated Data Augmentation for Precise Estimation of Organic Pollutants in Water with Deeplearning Neural Network

Conventional TOC estimation methods suffer from degradation, instability, and sensitivity to noise, resulting in substantial prediction errors. In this study, we propose a novel data augmentation approach, Variational Autoencoder-Regulated Data Augmentation (VADERA), which leverages a conditional variational autoencoder to generate synthetic spectral samples reflective of real ultraviolet-visible (UV-Vis) absorbance data. These augmented datasets address issues related to small sample sizes and overfitting, thereby improving model generalizability. When trained on VADERA-augmented data, a custom deep learning model (KimNet) exhibited rapid convergence and minimal validation loss, achieving high prediction fidelity. Latent space analysis revealed clear clustering aligned with total organic carbon (TOC) concentrations, validating the generative model’s structural integrity. Compared to classical and GAN-based augmentation strategies, VADERA significantly improved regression performance in terms of both R2 and RMSE. These findings demonstrate that VADERA provides a robust, accurate, and scalable solution for TOC estimation, offering a promising pathway toward cost-effective, AI-enhanced environmental monitoring.