Application of Large Language Models to Predict Enteric Viruses Removal Using Surrogate Indicators

Enteric virus monitoring in wastewater is constrained by sparse, costly data sets and imperfect surrogate relationships, limiting predictive modeling and facility-wide generalization. Here, we introduce CALM-VR (Constrained Augmentation with Large Language Models and Machine Learning for Viral Removal), a framework that couples schema-constrained synthetic data generation using GPT-5.2 with machine learning to predict enteric virus log removal. The study used 48 curated sampling data sets from five U.S. water reclamation facilities, comprising physicochemical variables, viral surrogates, and enteric viruses. Synthetic-data fidelity was supported by real versus-synthetic distributional agreement, no significant post-correction differences across 157 facility-parameter comparisons, preservation of pairwise dependence structure (mean absolute correlation difference = 0.151; root-mean-squared error (RMSE) = 0.218), and substantial overlap in principal component space. Under the primary removal-only scenario, tree-based models outperformed linear and latent-variable benchmarks for all four targets. When evaluated on real observations held out from model fitting, prediction from measured data alone was strongly limited by data scarcity (R² ≈ 0.11-0.46), and synthetic-data augmentation improved prediction of those held-out observations (R² ≈ 0.71-0.89). Although leave-one-location-out validation revealed reduced cross-facility transferability, CALM-VR provides a reproducible, quality-controlled framework for improving enteric virus removal prediction under data-scarce conditions and advancing wastewater surveillance.