For example, a train carrying dangerous goods that derailed in East Palestine, Ohio in 2023. In 2025, a series of devastating wildfires destroyed Los Angeles. In both cases, toxic plumes – clouds of harmful aerial materials that disperse time and space due to wind and turbulence were released.
St-Gasnet is trained with data from traditional computational fluid dynamics simulations that capture the movement of complex wind structures and plumes around urban buildings, roads and structures.
“St-Gasnet learns how plumes behave in urban areas from previous high-resolution simulations,” says LLNL scientist and author Giselle Fernández-Godino. “Looking at the first few minutes of the plume's release, we'll use these observations to predict how the plume will spread in the next few minutes.”
To predict how dangerous materials will move, the model learns the plume's velocity and acceleration patterns. It can also handle the discontinuity of the movement of the plume, such as when it hits a building and splits it into two. St-Gasnet works well without being told about the direction or speed of the wind. These conditions can be learned directly from the early actions of the plume.
Once trained, this model provides faster calculation speeds suitable for real-time emergency response.
“We hope this will support emergency responders and support their evacuation plans,” said author Yinan Wang, who was an intern at LLNL. “It acts as a component of an early warning system, allowing decision makers to take action more time and potentially integrates with mobile sensing and surveillance systems for real-time updates.”
Going forward, the team aims to develop a framework that can explicitly quantify uncertainty, and to estimate the potential impacts in each location, not just areas that have affected the urban plume. They are also looking for ways to optimize atmospheric sensor networks and calibration technologies to systematically align predictions and data.
Click here for more information.
