Evaluating CMAQ Representation of Wildfire PM₂.₅ in California (2008–2018): Regime-Dependent Performance and Nonlinear Error Structure

We evaluated daily paired Community Multiscale Air Quality (CMAQ) estimates and observations during California fire seasons (June–October), 2008–2018, using conventional metrics, classifications at 12 and 35 µg m⁻³, two case studies, and a leakage-resistant Random Forest diagnostic. After quarantining 18 CMAQ total PM₂.₅ values above 557 µg m⁻³ without capping, the primary analysis contained 187,051 records. Observed and modeled means were 10.69 and 9.29 µg m⁻³; Pearson r was 0.502 (r² = 0.252), mean bias was − 1.40 µg m⁻³, and RMSE was 9.96 µg m⁻³. CMAQ detected 45.4% of observed events at the descriptive/former annual level of 12 µg m⁻³ and 40.8% at the current 35 µg m⁻³ 24-hour standard level. At 35 µg m⁻³, CMAQ produced more events overall but missed 1,273 of 2,151 observed events because modeled and observed events often did not coincide by site-day. The Random Forest excluded observed PM₂.₅, CMAQ total PM₂.₅, and the no-fire component from its primary features. In an untouched site-year test partition (n = 37,430), R² was 0.711, RMSE was 5.82 µg m⁻³, and MAE was 3.44 µg m⁻³; fire-only CMAQ PM₂.₅ was the dominant predictor. CMAQ remains useful for regional and multi-year analyses, but paired-event detection and localized extreme-smoke estimates require explicit uncertainty treatment.