Accurate air quality forecasts are very important for keeping people healthy. However predicting pollution is often challenging because emissions can change from one place to another and over time, often in ways that current forecast models can’t capture. Machine learning models can help when frequently updated with new data, making predictions more accurate and better suited to local changes like weather and pollution sources.

In this study, we look at how to best retrain machine learning models that predict fine particles in the air (PM2.5). We use ML models and data from various cities around the world. The models are trained on past data covering 6 to 24 months, tested using different data sets, and then retrained every 6 or 12 months with new data. This process mimics real-life situations where models need to stay accurate as conditions change.

We combine local measurements with data from NASA’s air quality forecasts, and we use methods to understand factors affecting the predictions. Our results show that retraining models yearly leads to much better accuracy than updating less often, we demonstrate that updating models yearly is the best way to keep air quality forecasts reliable as cities and pollution change.