Varuna: Integrating AI with Environmental Data for Water Quality Predictions

Abstract

Water quality monitoring is essential for protecting human health and ecosystems, yet traditional analysis methods are too time-consuming to process effectively. To address this problem, we tested the hypothesis that an AI system trained on environmental data can accurately predict water quality conditions. Publicly available water quality sample data for Austin, Texas, was combined with environmental data from an API (Application Programming Interface) over an overlapping three-month period. Key parameters—including pH, dissolved oxygen, turbidity, nitrate, phosphorus, water temperature, max temperature, total precipitation, evaporation, and daylight hours—were processed and utilized to assign Water Quality Index categories. A Random Forest classifier was evaluated for accuracy and variable significance; using these results, a FLAN-T5 language model² was fine-tuned to generate both water-quality predictions and natural-language explanations from numeric inputs. The refined FLAN-T5 model² produced water-quality classifications along with coherent explanations aligned with these traits. An example result from Varuna is: "Water Quality Category: Fair. Water quality is primarily influenced by low dissolved oxygen, low nitrate levels, pH within optimal range, seasonal daylight effects." These findings indicate that AI systems can match traditional performance while improving usability. The model is presented as a user-friendly public website for everyone to predict future forecasts with up-to-date statistics.