Project: DL4MCS — Deep Learning Methods to Enhance Subseasonal Predictions of Mesoscale Convective Systems by Physics-based Systems

Investigators:

Dr. Hansi Singh, Planette AI (PI)

Dr. Susannah Burrows, Pacific Northwest National Laboratory (co-I)

Dr. Stefan Rahimi, University of Wyoming (co-I)

This project will use artificial intelligence and advanced Earth system models to make weeks-ahead forecasts of dangerous storm systems more accurate and useful for communities, businesses, and governments across the United States. These storms, called mesoscale convective systems (MCS), produce intense rain, hail, high winds, and sometimes tornadoes, and are a major cause of flooding, infrastructure damage, and economic disruption.

Why these storms matter

In many parts of the country, MCS supply a large share of the warm-season rainfall that feeds rivers, reservoirs, and groundwater, directly affecting water available for drinking, irrigation, and energy production. At the same time, they are responsible for more property losses than wildfires, hurricanes, and earthquakes combined, damaging homes, roads, power lines, and critical facilities. Better advance warning of where and when these systems are likely to strike can improve preparedness, reduce losses, and support safer, more efficient use of water and energy resources.

The forecasting gap

Traditional weather forecasts are very good for a few days in advance, but their skill drops sharply beyond about a week, especially for complex storm clusters like MCS. Longer-range “subseasonal” forecasts (7 days to 6 weeks) can capture broad patterns in temperature and moisture but often miss the fine-scale details that determine whether a community experiences a routine thunderstorm or a damaging multi-hour deluge. This gap leaves water managers, emergency planners, utilities, and insurers without the reliable, location-specific guidance they need to plan for high-impact events weeks ahead.

What this project will do

Our funded project will build a new forecasting system that combines the strengths of physics-based climate and weather models with modern AI methods to improve storm predictions on subseasonal timescales. First, we will use AI to “expand” existing forecast ensembles, effectively generating many more plausible future scenarios so that rare but dangerous events are better represented. Second, we will train AI models to blend forecasts with key observations of the ocean and land surface—such as soil moisture and upper-ocean heat content—because these slowly changing conditions strongly influence whether environments become favorable for MCS.

High-resolution storm detail

Finally, we will apply advanced generative AI techniques to translate coarse, global forecasts into high-resolution maps, detailed enough to represent individual storm systems and their evolution over the central and eastern United States. These methods are trained using state-of-the-art regional simulations and cloud observations, allowing the system to better capture how storm structure, rainfall intensity, and cloud processes respond to large-scale atmospheric patterns. By comparing different plausible representations of cloud physics, we will also quantify uncertainty in rainfall predictions and identify configurations that match observations best.

Benefits to society and the energy–water system

The end result will be a prototype forecast pipeline that can predict MCS activity over the U.S. up to several weeks in advance with substantially higher skill than today’s operational models. These improved forecasts can inform reservoir operations, drought and flood preparedness, grid reliability planning, and risk management decisions in sectors such as agriculture, transportation, and insurance. By directly supporting the Department of Energy’s goal of “Predicting U.S. Water for Energy: Weeks to Years,” the project will help communities and industries build resilience to extreme storms while making smarter, more sustainable use of water and energy.