As wildfire smoke becomes one of the most dangerous and costly hazards in the U.S. and globally, better predictions could mean earlier warnings and healthier communities. When fires burn, their smoke plumes release massive amounts of aerosol pollution. These tiny particles interact with sunlight and clouds, influencing weather, air quality, and Earth systems. Wildfire emissions are challenging to predict, but they are a crucial component in atmospheric composition models, supporting air-quality forecasting and our understanding of aerosol impacts on weather and climate. 

Researchers from CIRES, NOAA’s Global Systems Laboratory, George Mason University, and UT Arlington are tackling this challenge head-on. Today at the 2026 American Meteorological Society annual meeting, the team will present a new AI-driven system. The goal is to predict wildfire emissions weeks in advance, a breakthrough that could transform sub-seasonal to seasonal (S2S) weather forecasts (35 to 45 days in advance). 

Today, many weather models use persistence — a forecasting concept that assumes the current day’s fire emissions will continue tomorrow. While this assumption introduces uncertainty even in short-range forecasts (about five to seven days), its impact is far more pronounced for S2S prediction. As a result, some S2S models use simplified or historical-average fire emissions, which fail to capture the day-to-day variability and extremes of real wildfire seasons, reducing forecast accuracy.

The challenge

Wildfire emission inventories from different agencies often disagree, causing large uncertainties in aerosol and atmospheric composition forecasts. There's no single, continuous record of fire emissions with a wide geographic spread, so scientists typically evaluate fire emission performance indirectly by comparing model outputs with observations.

The solution

The team developed an AI-driven system that combines data from 7 global fire emission inventories, including major fire-emitted aerosols or their precursors, such as organic carbon, black carbon, and sulfur dioxide. The model uses meteorological data, vegetation condition, land use, and fire radiative power — a measure of fire intensity — as inputs. Trained on four years of data, the system forecasts fire radiative power and then converts it into predicted emissions.

Testing the Machine Learning/AI model

Researchers tested the system during two key periods. At the weather-scale, real-time seven-day predictions from October 1, 2025, to present show that Machine Learning/AI-based fire-emission estimates are promising for short- to medium-range and S2S prediction. At the S2S scale, they tested an average fire season and an extreme wildfire season and found limited skill in simulating large wildfires.

What’s next

The team is working to optimize machine learning/AI algorithms and develop post-processing correction methods to improve accuracy, refine approaches for estimating fire spread and ignition probability, and integrate these forecasts into real-time or operational model systems. Wildfires are no longer rare events; this research could be a game-changer for early prediction of air quality and climate impacts from wildfire smoke.