Integrating Atmospheric Physics and Spatiotemporal Machine Learning to Project Wildfire Smoke Exposure and Health Risks in Canada under Future Climate

The catastrophic Canadian wildfire seasons of 2024-2025, including the devastating Jasper fire, underscore the escalating public health risk from wildfire smoke. Toronto and other Eastern Canadian cities repeatedly experienced the world's worst air quality during prolonged smoke episodes. Using machine learning, we project future wildfire activity and associated smoke exposure under SSP585 by deriving a joint spatiotemporal distribution of wildfire ignition, size, and smoke dispersion that explicitly accounts for regional correlation dependencies. Our framework integrates the Canadian National Fire Database, CMIP6-derived vapor pressure deficit (VPD), and quantile regression to distinguish the drivers of small, wind-driven events versus large, climate-driven fires. Preliminary results show a 35%-85% increase in population exposure to hazardous smoke (PM > 50 μg/m³) by mid-century, with Eastern Canada and U.S. cities most affected. These findings highlight urgent mitigation and adaptation needs.