Federal scientists announced a new artificial intelligence tool that can forecast drought conditions 90 days ahead across the entire United States, marking a significant step in how the government predicts and communicates water scarcity risks. The system, built on machine learning models trained to predict streamflow at thousands of monitoring stations, arrives as climate change intensifies the frequency and severity of drought events that cost American agriculture billions of dollars each year.
A 90-Day Window Into Drought Risk
The U.S. Geological Survey recently unveiled a nationwide AI drought forecasting system designed to give water managers and farmers a three-month lead time on emerging dry spells. The tool is powered by machine learning models that generate weekly streamflow percentile forecasts one to 13 weeks into the future, covering more than 3,000 USGS gages across the contiguous United States (according to the agency’s published model outputs).
Two algorithms drive the predictions: Long Short-Term Memory (LSTM) networks, which excel at capturing time-dependent patterns in river flow data, and LightGBM, a gradient-boosting framework suited to tabular environmental inputs. Both outperformed standard benchmarks, including simple persistence and ARIMA statistical models, in testing. The combination matters because streamflow is the variable most directly tied to irrigation supply, reservoir operations, hydropower generation, and municipal water planning. A forecast that can reliably flag low-flow conditions weeks in advance gives decision-makers time to adjust reservoir releases, coordinate voluntary conservation measures, or shift planting schedules before shortages become acute.
USGS scientists emphasize that the tool is not meant to replace existing drought indicators such as soil moisture or precipitation anomalies, but to complement them with a direct measure of water availability in rivers. By expressing predictions as percentiles relative to historical flows, the system also allows local water managers to interpret forecasts in the context of their own experience with past dry years.
NOAA’s AI Ensemble Fills a Speed Gap
The USGS initiative fits into a broader federal push to incorporate artificial intelligence into environmental prediction. NOAA has been building its own AI-based weather and climate infrastructure since late 2025, when the agency operationally deployed a new generation of global models that use machine learning surrogates alongside traditional physics. That rollout included the AI Global Ensemble Forecast System (AIGEFS) and the Hybrid-GEFS, described in NOAA’s program overview of AI-driven models developed with external partners.
Early evaluations from NOAA’s Global Systems Laboratory indicate that AIGEFS improves skill compared with the legacy Global Ensemble Forecast System at key lead times, particularly for large-scale circulation patterns and precipitation anomalies, according to internal testing summarized by the lab’s research updates. Crucially, the new AI models run much faster than full-physics systems, enabling more frequent and larger ensembles that better sample uncertainty.
These AI-generated forecasts are distributed through NOAA’s Operational Model Archive and Distribution System, known as NOMADS, which provides public access to gridded weather and climate data; the drought community can ingest these archived fields to drive hydrologic and soil moisture models. The speed and accessibility of AI ensembles make it easier to update seasonal outlooks more often, test alternative scenarios, and drive downstream tools like the new USGS streamflow predictor without overwhelming computing resources.
Why Seasonal Drought Prediction Has Been So Hard
The Climate Prediction Center, the federal office responsible for issuing official seasonal drought outlooks, has long relied on a blend of physical drivers to build its assessments. Those drivers include current snowpack levels, the timing of spring melt, temperature and precipitation outlooks, and the state of the El Niño/Southern Oscillation cycle, as described in the center’s expert guidance. Human forecasters weigh these ingredients alongside historical analogs to judge where drought is most likely to develop, persist, or ease over the coming months.
The difficulty is that these inputs interact in complex ways that are hard for traditional statistical techniques to capture, especially at the one-to-three-month horizon where deterministic weather forecasts lose skill but long-term climate averages are too coarse to guide local water decisions. This “subseasonal to seasonal” window has been a persistent weak spot in the forecast chain, leaving farmers and reservoir operators to make high-stakes choices with limited predictive information.
NOAA’s Weather Program Office has been exploring AI and machine learning methods to bridge this gap, focusing on models that can learn subtle patterns in high-dimensional environmental datasets. Unlike linear regression or simple analog schemes, neural networks and ensemble methods can ingest dozens of predictors, sea-surface temperatures, soil moisture, atmospheric circulation indices, and more, and discover nonlinear relationships that matter for drought onset and persistence. The USGS streamflow tool is one manifestation of this broader shift toward data-driven approaches that complement physically based models rather than replace them.
Peer-Reviewed Evidence Backs the Approach
The scientific case for AI-driven seasonal forecasting has strengthened rapidly over the past few years. One influential study in Communications Earth and Environment showed that machine learning algorithms trained on output from the Community Earth System Model Large Ensemble can produce skillful seasonal precipitation predictions by learning relationships between large-scale climate patterns and regional rainfall. By operating on climate model simulations instead of purely observational records, the system was able to exploit a much larger sample of possible climate states, improving robustness.
Another line of research has focused on AI models that operate more like traditional numerical weather prediction systems but with learned dynamics. A study in npj Climate and Atmospheric Science evaluated the ACE2 machine learning weather model and found that it generates skillful one-to-three-month forecasts when initialized with reanalysis data, with performance quantified using metrics such as North Atlantic Oscillation correlation and spatial anomaly patterns. Those results suggest that AI can capture key modes of climate variability that strongly influence drought risk over North America and Europe.
Hybrid approaches are also gaining traction in the drought research community. For example, some teams have combined feature selection methods such as Random Forests with sequence models like LSTMs, allowing the first stage to identify the most informative predictors before the second stage learns temporal dependencies. Studies using this architecture report improved accuracy for indices such as the Standardized Precipitation Evapotranspiration Index, particularly at lead times beyond one month, compared with either method alone.
From Research to Operations
Turning experimental AI models into operational tools is not straightforward. Agencies must validate performance across diverse climate regimes, guard against overfitting, and ensure that systems remain stable when confronted with extremes outside the historical record. For drought forecasting, another challenge is communicating uncertainty: even a skillful model will sometimes miss the timing or intensity of dry spells, and water managers need probabilistic information rather than binary yes-or-no answers.
The USGS streamflow tool addresses some of these concerns by issuing percentile-based forecasts for thousands of individual gages, which can be aggregated into basin-level indicators or interpreted locally. NOAA’s AI ensembles, meanwhile, provide a natural framework for expressing uncertainty through spread among many slightly different simulations. Together, they allow forecasters to say not just whether drought is more likely, but how confident they are and which regions face the greatest risk.
Integration with existing decision-support systems will determine how much real-world impact these advances have. State and regional drought task forces, irrigation districts, and power utilities already rely on a patchwork of forecasts, from short-term weather outlooks to seasonal climate projections. Embedding AI-driven streamflow and precipitation predictions into those workflows, while preserving human oversight, could help operators stage water transfers earlier, fine-tune reservoir rule curves, or adjust planting recommendations in time for farmers to respond.
Scientists and agency officials caution that AI is not a silver bullet. Physical understanding of the climate system remains essential for diagnosing model failures, designing robust adaptation strategies, and anticipating how a warming world might alter the relationships that current algorithms have learned. But as droughts grow more frequent and costly, the combination of physics-based models, expert judgment, and carefully validated machine learning systems may offer the best chance of giving communities the lead time they need to prepare.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
