Old-school gradient boosting just beat three flavors of neural networks at forecasting one of the country's most unpredictable aquifers.
Researchers tested five model families: extreme gradient boosting (XGBoost), extremely randomized trees, LSTM, CNN, and Transformers, forecasting spring discharge and groundwater levels for the Edwards Aquifer in Texas using about 79 years of hydroclimatic data. XGBoost came out on top across every horizon, posting R2 scores of at least 0.97, 0.96, and 0.94 for the 1-4, 5-8, and 9-12 week forecast windows, and correctly flagging drought severity more than 90% of the time at the first three alert thresholds. The deep learning models kept pace at short range but degraded the further out they predicted, with occasional outright failures at longer horizons. The researchers then wrapped the winning model in a five-agent system that automates data collection, model selection, prediction, threshold monitoring, and reporting.
That result matters because karst aquifers are supposed to be the hard case for forecasting. Their recharge is nonlinear, triggered by individual storm events, and routed through wildly uneven underground flow paths, which is exactly the kind of messy, irregular data neural networks are usually pitched as better at handling. The researchers attribute the gap to tree-based splits mapping cleanly onto a handful of simple weather and recharge variables, while neural nets tend to smooth over the sharp spikes that define an actual flood or drought.
Notably, the paper's 'agentic AI harness' framing describes a system built around a model the authors themselves call deterministic, a reminder that 'agentic' increasingly gets attached to what is, underneath, a fixed automation pipeline.