Setting up computer models that predict water flow in rivers and streams requires hundreds of complex decisions about how to represent processes like snowmelt, evaporation, and groundwater flow. Current automated methods can only adjust numbers, not make strategic decisions about which processes to include or how to represent them.

We built an AI assistant that learns to configure water models like an expert hydrologist. The AI runs experiments, evaluates results against multiple objectives (streamflow accuracy, snow dynamics, water storage), and learns from both successes and failures. It maintains memory of what works for different watershed types and references scientific literature to guide decisions. Unlike traditional optimization that just tweaks parameters, our AI explains its reasoning—like 'trying a different snowmelt equation because spring peaks aren't captured well.'

We're investigating fundamental questions: Does AI develop scientific intuition? Does it recognize that desert watersheds behave differently than mountainous? How much should it remember from previous experiments?

This could democratize advanced water modeling globally, especially in regions with limited technical expertise. It might also discover novel insights by systematically exploring model configurations that humans haven't considered, potentially accelerating our understanding of water resources in a changing climate.