Predicting where wildfire smoke will go next

By Emma Frankham

In recent summers, wildfire smoke has been inescapable across Wisconsin, drifting south from Canada and leading to long stretches of poor air quality. Whether that smoke clears or lingers is often influenced in part by the atmospheric boundary layer - the lowest part of the atmosphere, where airborne particles mix and move.

At night or during cooler weather, the boundary layer compresses toward the ground, sometimes to heights below 100 meters above the surface. Under these conditions, pollutants, including smoke, can become trapped near the surface, increasing exposure and health risks. During the day, as the ground warms, the boundary layer rises and can reach heights of 2,000 meters, enhancing vertical mixing and the dispersion of pollutants. Tracking these changes in near real time is critical for predicting how pollution levels shift and how they may affect public health.

In a new study, University of Wisconsin-Madison statistics professor Chris Geoga and atmospheric scientist Paytsar Muradyan of Argonne National Laboratory are tackling this challenge using machine learning and high-resolution data. Using data from the Department of Energy's CROCUS Urban Integrated Field Laboratory - a large-scale collaboration led by Argonne - the team developed a model that can estimate boundary layer height in near real time. The work is an important step toward helping public health officials better anticipate how air quality conditions, including smoke impacts, may change - and provide earlier warnings to the public.

The challenge of tracking a fast-changing boundary layer

Doppler LiDAR sends pulses of light into the atmosphere and measures how they scatter off particles, providing information used to infer boundary layer structure and height. But the data is complex and difficult to interpret. Clouds, rainfall, and even insect swarms can contaminate the signal. "Insect swarms can produce absolutely outrageous data," reflects Geoga. "There's a lot of information hidden in the layers of the raw data that we're not seeing," adds Muradyan.

This uncertainty means the diagnosed height of the boundary layer is often misidentified, and with a significant time lag. "Estimating boundary layer height isn't usually done at very high frequency," explains Muradyan. "Often you need to average the data to improve the signal quality, so you might only get an estimate every 30 minutes or every hour - but many atmospheric processes happen on much shorter timescales - minutes, not hours. And that's where we start to lose important detail." These estimates can miss rapid fluctuations in the boundary layer, such as during the morning when it begins to rapidly deepen, or from more localized phenomena such as lake breezes or urban heat.

Applying statistical methods to an atmospheric problem

To address these gaps, Geoga turned to statistical modeling. Traditional meteorological approaches often rely on simplified assumptions that fail to account for boundary layer height estimates across space and time. For example, a boundary layer height estimated at 2 p.m. is often treated as independent from one taken just 30 minutes later. In reality, the underlying measurements, which are essentially snapshots of vertical wind motion, are strongly linked from one moment to the next and indirectly related to the boundary layer height.

Leveraging this dependence, Geoga developed a machine learning model that treats the boundary layer as something that evolves over time, revealing "beautiful fluctuations" in the data. By taking into account relationships between successive measurements of vertical wind, the approach can track both gradual shifts and sudden changes, capturing patterns that would otherwise be smoothed away.

Instead of analyzing each moment in isolation, the model learns from how conditions change over time to infer boundary layer height. And because this approach doesn't treat these measurements as independent of each other, another advantage is that it reduces the time lag for identifying boundary layer changes from hours to minutes; a shift that dramatically improves how quickly scientists can detect meaningful changes. The approach, Geoga explains, "is far more statistically and computationally demanding than it might seem."

Planes, pollen, and pollution

More accurate tracking of the boundary layer could improve air quality forecasts and support earlier warnings during events such as wildfire smoke episodes. In practice, this means communities could receive more timely alerts about dangerous air conditions. And because its height is a critical parameter in weather forecasting, it can also help researchers and meteorologists better anticipate a host of other events influenced by the boundary layer, including aircraft turbulence and the spread of airborne particles like pollen.

The work illustrates how statistics and data science can unlock new insights in complex, real-world systems spanning atmospheric science, meteorology, and public health. By making boundary layer changes visible in near real time, the research offers a powerful new tool for understanding, and responding to, air quality challenges. Georga's work also reflects a broader approach in the newly launched College of Computing & Artificial Intelligence, where computational methods and domain expertise come together to address challenges in environmental monitoring, forecasting, and public health.

Related: