Evaluating Chemical\nTransport and Machine Learning\nModels for Wildfire Smoke PM<sub>2.5</sub>: Implications for Assessment\nof Health Impacts

Growing wildfire smoke represents a substantial threat\nto air quality\nand human health. However, the impact of wildfire smoke on human health\nremains imprecisely understood due to uncertainties in both the measurement\nof exposure of population to wildfire smoke and dose–response\nfunctions linking exposure to health. Here, we compare daily wildfire\nsmoke-related surface fine particulate matter (PM<sub>2.5</sub>) concentrations\nestimated using three approaches, including two chemical transport\nmodels (CTMs): GEOS-Chem and the Community Multiscale Air Quality\n(CMAQ) and one machine learning (ML) model over the contiguous US\nin 2020, a historically active fire year. In the western US, compared\nagainst surface PM<sub>2.5</sub> measurements from the US Environmental\nProtection Agency (EPA) and PurpleAir sensors, we find that CTMs overestimate\nPM<sub>2.5</sub> concentrations during extreme smoke episodes by up\nto 3–5 fold, while ML estimates are largely consistent with\nsurface measurements. However, in the eastern US, where smoke levels\nwere much lower in 2020, CTMs show modestly better agreement with\nsurface measurements. We develop a calibration framework that integrates\nCTM- and ML-based approaches to yield estimates of smoke PM<sub>2.5</sub> concentrations that outperform individual approach. When combining\nthe estimated smoke PM<sub>2.5</sub> concentrations with county-level\nmortality rates, we find consistent effects of low-level smoke on\nmortality but large discrepancies in effects of high-level smoke exposure\nacross different methods. Our research highlights the differences\nacross estimation methods for understanding the health impacts of\nwildfire smoke and demonstrates the importance of bench-marking estimates\nwith available surface measurements.