Initiative supported by the U.S. Department of Energy will test deep learning with physical models to anticipate storms and extend preparation time for communities, utilities, and insurers in the face of strong winds, rain, and floods.
A project supported by the U.S. Department of Energy will test artificial intelligence in predicting large-scale storms between seven days and six weeks before their formation in the country’s continental territory.
Storm prediction still faces a one-week limit
According to interestingengineering, Planette AI, the Pacific Northwest National Laboratory, known as PNNL, and the University of Wyoming are working together on the DL4MCS project, part of Phase I of the Genesis Mission program of the United States Department of Energy.
The initiative seeks to improve the prediction of mesoscale convective systems, identified by the acronym MCS in English. These systems are large clusters of storms capable of occupying areas with hundreds of kilometers in extent.
-
A giant of almost 5 tons abandoned by SpaceX is traveling through space at 8,700 km/h and will crash onto the Moon, causing an explosion equivalent to 3 tons of TNT, creating a new crater, and launching debris that could be observed from Earth.
-
Refrigerator with days numbered: Xiaomi creates a smart 621-liter version that works with Alexa, makes only 35 dB of noise, freezes 14 kg in 24 hours, and arrives with a surprising price in Europe
-
Harvesting 6.62 million tons and declining by 16% in the 2025 season, Brazil’s wheat production shrank contrary to the record grain harvest, in a cycle where the country remained dependent on imports.
-
At 16 years old, a student creates a system with mirrors that tracks the Sun to heat weeds, reduce herbicides, and wins an international environmental engineering award.
The phenomena can cause strong winds, heavy rains, floods, and other severe weather conditions. They also account for a significant portion of the precipitation recorded during the warm season in the United States.
Currently, accurately predicting the activity of these systems more than a week in advance remains a challenge. The difficulty lies in the different scales involved in the formation and development of the storms.
The atmospheric conditions responsible for the formation of the systems can evolve over extended periods. However, the storms are driven by processes that occur on much smaller scales, making their representation by existing prediction systems difficult.
Deep learning will be combined with physical models
DL4MCS will investigate predictions between approximately seven days and six weeks, an interval known as sub-seasonal forecasting. The project will assess whether deep learning can help fill the existing gap in this period.
The proposal is not to replace weather models based on physical principles. Researchers will test whether computational methods can extract useful patterns from available forecasts and better indicate when and where systems are likely to develop.
The team will analyze atmospheric conditions on different scales. The work will cover everything from broad climate patterns capable of influencing storm formation to cloud-scale processes that determine how these events evolve.
According to Hansi Singh, founder and CEO of Planette AI, the combination of artificial intelligence and proven physical systems aims to make advance information about storm risks more useful for sectors and communities dependent on better forecasts.
PNNL will contribute expertise in Earth system modeling and atmospheric process assessment. The University of Wyoming will focus on regional modeling and downscaling forecasts.
Regional models will attempt to produce more accurate guidance
Large-scale weather models typically work with resolutions considered too coarse to represent all the details necessary for local storm forecasting. Downscaling seeks to convert this information into higher resolution and regionally relevant data.
Susannah Burrows, an atmospheric scientist at PNNL, stated that improving forecasts requires advances from large-scale climate factors to the cloud microphysics responsible for shaping storm behavior.
According to Burrows, the collaboration brings together expertise in Earth system modeling, artificial intelligence, and process-level model assessment to explore a new path in sub-seasonal forecasting.
Stefan Rahimi, a professor at the University of Wyoming, declared that transforming large-scale, low-resolution forecasts into more precise and decision-relevant guidance is essential for improving preparedness and resilience.
Advance warnings could support public services and communities
If the approach works, an early indication of high risk could grant more time for utility companies to prepare for potential disruptions.
The information could also help insurers assess potential exposures and enable communities to plan for severe storms and flooding. These benefits are possibilities conditioned on the success of the investigated approach.
The project is part of a broader effort by the Department of Energy to apply advanced computing and machine learning to scientific problems still difficult to solve by conventional methods.
The Genesis Mission brings together researchers from government, industry, and academia to develop approaches in energy, science, and national security. At DL4MCS, the immediate goal is to verify if physical forecasts and deep learning can anticipate large storm systems.
