“The goal in accurately predicting air pollution caused by wildfires is early warning—providing enough time to public health officials and emergency managers to assess the impact of the air quality on a particular community. Our existing early warning systems are built quite well for Western wildfires, from which smokes are usually transported to the East Coast in June or July. What happened this June was an Eastern wildfire that was transported downward in a swirl.

“Given the effect of this Eastern wildfire, it’s clear we need better models—rare-case models that react to triggers, or extreme cases. Traditional forecasting models are only trained for baseline, no-smoke, or smoke scenarios. We found it best to merge these two into our ST-Transformer model, which also determines other important factors associated with wildfires: meteorological factors, pollution factors like traffic, local vegetation, and fuel. We call this the sparse attention mechanism, which means the model brings together different variables that contribute to wildfires and impact their smoke. Sparse attention can also better estimate PM2.5—the fine particulate matter in wildfire smoke that poses risks to health—because it reduces overestimations during baseline times and underestimations that traditional models make, since they are using historical data. Our model can better capture the changes in PM2.5 when there’s a wildfire, while not underestimating the amount of PM2.5 that is already present in the air, and we can train it to better capture rare cases.

“The greatest issue has to do with health. We study PM2.5, but smoke can also contain soot, black carbon, particles of heavy metal. Fuel and moisture in a region also play a role. If we can warn people earlier about hazardous air quality, we can address both acute and chronic effects of wildfire smoke on health.”