In 2023, a train carrying hazardous materials derailed in East Palestine, Ohio. In 2025, a series of destructive wildfires ravaged Los Angeles. In both cases, a toxic plume - a cloud of harmful airborne materials that disperse over time and space due to wind and turbulence - was released.
Toxic plumes from industrial accidents, chemical spills and structural fires can pose immediate and serious health and environmental risks, particularly in densely populated urban areas. Existing computer models to predict how plumes travel can take hours to run, leaving emergency responders without fast and reliable predictions and hindering evacuation planning and early-warning systems.
In a study published in PNAS Nexus, researchers from Lawrence Livermore National Laboratory (LLNL) described a new deep learning model, called ST-GasNet, capable of predicting toxic plume behavior in just a few minutes.
ST-GasNet is trained with the data from traditional computational fluid dynamics simulations, which capture the complex wind structures and plume movement around buildings, streets and structures in urban areas.
"ST-GasNet learns how plumes behave in urban areas from previous high-resolution simulations," said LLNL scientist and author Giselle Fernández-Godino. "It looks at the first few minutes of a plume release and uses those observations to predict how the plume will spread in the next several minutes."
To predict how hazardous material moves, the model learns the velocity and acceleration patterns of the plume. It can also handle discontinuities in plume movement, like when a plume hits a building and splits in two. ST-GasNet even performs well without being told the wind direction or speed: it can learn those conditions directly from the plume's early behavior.
Once trained, the model offers faster computational speeds that are feasible for real-time emergency response.
"We hope this will support emergency responders and help with evacuation planning," said author Yinan Wang, who was an intern at LLNL. "It can serve as a component in early warning systems, giving decision-makers more time to act, and potentially integrate with mobile sensing and monitoring systems for real-time updates."
Looking ahead, the team aims to develop a framework that can explicitly quantify uncertainty, estimating not only the plume-impacted area in a city but the likelihood of impact at each location. They are also exploring ways to optimize atmospheric sensor networks and calibration techniques to systematically align predictions and data.
