Explainable Machine Learning for Air Quality and Meteorological Prediction: From Intrinsic Design to Post-Hoc Validation

Machine learning models are increasingly used to forecast air pollutant concentrations (e.g., PM2.5, PM10, NO2, O3) and related meteorological variables (temperature, humidity, wind speed and direction, atmospheric pressure), supporting public health alerts, regulatory compliance, and environmental policy. These models are often high-performing but opaque, which is a particular concern in this domain: pollutant and meteorological variables are strongly correlated and spatiotemporally structured, sensor data (especially from low-cost monitoring networks) is frequently noisy or missing, and model outputs can directly inform public health decisions and regulatory thresholds. Explainable AI (XAI) offers tools to address this opacity, but intrinsic and post-hoc approaches are rarely evaluated against one another in this domain, and standard explanation methods are not always validated against the specific statistical structure of environmental data.This Research Topic aims to advance the explainability of machine learning models used for air quality and meteorological prediction, with three specific aims. First, to examine how intrinsic (e.g., physically constrained, sparse, or rule-based) and post-hoc explanation approaches can be integrated across the modeling pipeline for pollutant and meteorological forecasting, rather than applied in isolation. Second, to strengthen the evaluation of explanations in this domain specifically, addressing faithfulness to model reasoning, robustness and stability under sensor noise or missing data, consistency across correlated meteorological drivers, calibrated uncertainty (particularly relevant given sensor and forecast uncertainty), and interpretability for the domain experts, regulators, and public stakeholders who act on these predictions. Third, to clarify conceptual boundaries relevant to deployment: explainability and interpretability describe properties of a model's output, while regulatory trust and public accountability are broader goals that explanation alone does not guarantee. Submissions applying standard explainability tools to air quality or meteorological data without addressing the domain's specific statistical or decision-making challenges are not a good fit for this Topic.We welcome Original Research, Methods, Reviews, and Perspective articles on:- Interpretable-by-design models for pollutant and meteorological forecasting, including physically informed, sparse, or rule-based architectures, and rigorous assessment of what these designs do and do not guarantee about interpretability.-Post-hoc explanation adapted to spatiotemporal environmental data, including methods that account for spatial autocorrelation, temporal dependence, and confounding between meteorological and pollutant variables.- Evaluating explanation quality under real-world sensor conditions, including robustness to missing or noisy low-cost sensor data, faithfulness, consistency, and uncertainty quantification.- Human-centered explainability for decision-makers, including studies with public health officials, regulators, or affected communities on how explanations support (or fail to support) real decisions.-Cross-domain integration, such as joint modeling of meteorological drivers and pollutant dynamics, where explainability clarifies causal versus correlational relationships between variables. Topic editors: Nilesh Maltare, Safvan Vahora, Ashwini Kumar Jha Source: https://www.frontiersin.org/research-topics/83693/explainable-machine-learning-for-air-quality-and-meteorological-prediction-from-intrinsic-design-to-post-hoc-validation