Air Quality Arena: New benchmark tests foundation models across 6 pollutants in 7 countries

A new paper from Rishi Bharadwaj, Manik Gupta, and Pandarasamy Arjunan introduces Air Quality Arena (AQA), a comprehensive dataset and benchmark designed to evaluate time-series foundation models (TSFMs) for short-term air quality forecasting. AQA covers six major pollutants—PM2.5, PM10, O3, NO2, SO2, and CO—over a three-year period across seven countries spanning four continents: India, China, the US, the UK, South Korea, Australia, and South Africa. The dataset includes over 14,000 station-pollutant series, making it one of the most geographically and chemically diverse benchmarks for air quality tasks. The authors benchmarked 11 leading TSFMs alongside classical baselines, finding that TSFMs consistently outperform traditional methods in zero-shot forecasting. The top-performing model uses a cross-modal architecture that adapts a vision foundation model for time series data, highlighting the potential of transfer learning across domains.

Air pollution is estimated to cause 7.9 million premature deaths annually, making accurate forecasting a critical public health priority. Existing benchmarks often lack geographic breadth and pollutant coverage, and they fail to evaluate the latest generation of foundation models. AQA fills this gap by providing large-scale, real-world data and a standardized evaluation framework. The dataset and code are publicly released, enabling researchers and practitioners to test new models against a consistent baseline. This work underscores the promise of foundation models for environmental applications, where data scarcity and variability across regions have historically limited model generalization.