Machine Learning Bias Correction of HRRR-Smoke Surface PM 2.5 Forecasts during the Alaska Wildfire Season

Smoke-related PM2.5 is the primary air quality concern during the summer in Alaska, yet accurate forecasting remains a major challenge. In this study, we use machine learning (ML) techniques to improve smoke forecasts from NOAA’s HRRR-Smoke model. We find that the model underestimates surface PM2.5 by a factor of up to five during the wildfire season in Alaska. We evaluated Random Forest (RF), one-dimensional, and two-dimensional convolutional neural network (CNN1D and CNN2D) models. Among them, CNN1D performed the best, reducing the underestimation factor to two or less. Analysis of the relationships between key predictors, such as Surface and Vertical Smoke and calibrated PurpleAir observations, suggests that errors in the vertical distribution of smoke are a primary source of underestimating bias. Atmospheric sounding data further show that the HRRR-Smoke model fails to capture daytime temperature inversion layers during wildfire events. This bias is likely caused by missing fire radiative power (FRP) detections under heavy smoke or cloudy conditions, which leads to low smoke concentrations and under-represented radiation feedback necessary to maintain near-surface inversion. Although accurately representing physical and chemical processes in models remains highly challenging, our results demonstrate ML offers an effective approach to improving daily surface PM2.5 forecasts in wildfire-prone regions like Alaska.