Predicting large thunderstorms weeks in advance remains a major challenge for weather forecasters. A new project supported by the U.S. Department of Energy is currently examining whether combining deep learning with traditional weather models can provide early warning of these powerful storm systems.
Planette AI, Pacific Northwest National Laboratory (PNNL), and the University of Wyoming are collaborating on a project called DL4MCS. This effort is part of DOE’s Genesis Mission Phase I program, which focuses on forecasting mesoscale convective systems (MCS) across the continental United States.
An MCS event is a collection of large thunderstorms that can spread over hundreds of miles. It can cause harmful winds, heavy rain, flooding, and other severe weather. It also accounts for a significant portion of the warm season rainfall in the United States.
The problem is that while the atmospheric conditions that cause these systems can change over long periods of time, the storms themselves are caused by processes that operate on a much smaller scale. This makes it difficult for existing prediction systems to accurately predict MCS activity beyond about a week.
Bridging the weather scale
DL4MCS investigates whether deep learning, in conjunction with physically-based predictive models, can help close that gap. The researchers target forecast windows ranging from approximately seven days to six weeks, known as subseasonal forecasts.
The idea is not to replace established weather models with machine learning systems. Instead, this project will test whether computational methods can extract useful patterns from existing forecasts and improve predictions about when and where MCS events are likely to occur.
The research team will examine atmospheric conditions at a variety of scales, from the broad climate patterns that can influence storm formation to the cloud-scale processes that determine how storms evolve.
“DL4MCS reflects Planette AI’s commitment to providing more actionable environmental intelligence for high-stakes decision-making,” said Founder and CEO Dr. Hansi Singh. “By combining cutting-edge AI with proven physical forecasting systems, this project aims to make storm risk information weeks in advance more useful to sectors and communities that rely on greater foresight.”
PNNL brings expertise in Earth system modeling and atmospheric process assessment, while the University of Wyoming will focus on regional modeling and downscaling. This is important because large-scale weather models typically have too coarse a resolution to capture all the detail needed to predict local storms.
“Improving predictions of mesoscale convective systems requires advances across scales, from large-scale climate factors to the cloud microphysics that shape storm behavior,” said Dr. Susannah Burrows, an atmospheric scientist at Pacific Northwest National Laboratory. “This collaboration brings together complementary strengths in Earth system modeling, AI, and process-level model evaluation to explore new paths toward better subseasonal forecasts.”
turn a forecast into a warning
The University of Wyoming’s contributions include converting large-scale forecasts into higher-resolution information that is more relevant at the regional level.
“The University of Wyoming is excited to contribute our regional downscaling expertise and responsible integration with AI forecasting to this effort,” said Dr. Stephen Rahimi, University of Wyoming Derecho Professor. “The ability to translate rough, large-scale predictions into higher-resolution, decision-relevant guidance is essential to improving real-world preparedness and resilience.”
If this approach works, its benefits could extend beyond simply predicting storms more accurately. Early indications of increased storm risk could give utilities more time to prepare for disruptions, help insurance companies assess potential risks, and allow communities to plan for severe weather and flooding.
The project is also part of a larger DOE effort to leverage advanced computing and machine learning to address scientific problems that remain difficult to solve using traditional methods. The Genesis mission brings together researchers from government, industry, and academia to develop new approaches across fields including energy, science, and national security.
The immediate goal of DL4MCS is a narrower but potentially important goal. The goal is to determine whether better use of existing physical forecasts and deep learning can make major thunderstorm systems more predictable weeks in advance of their occurrence.
This project is supported by the U.S. Department of Energy’s Genesis Mission Phase I program.
