Artificial intelligence for atmospheric chemistry: from observation to process understanding

Understanding the sources and evolution of air pollutants is difficult because emissions, meteorology, transport, and atmospheric chemistry interact across different spatial and temporal scales. Chemical transport models provide process-based descriptions but are computationally demanding, while receptor models infer source contributions from observations and depend on simplifying assumptions. Artificial intelligence (AI) provides new tools for combining atmospheric observations with model information. Here, we review AI applications in atmospheric composition, with a focus on spatiotemporal reconstruction, process interpretation, source attribution, and the effects of meteorology and emissions. We summarize the use of explainable AI and causal inference to examine relationships in atmospheric data, while emphasizing that statistical associations do not by themselves demonstrate physical or chemical mechanisms. We also outline an evidence-based approach for evaluating AI results using atmospheric knowledge, uncertainty, independent observations, experiments, and numerical models, and discuss emerging applications of physical constraints and foundation models.