Toward causal artificial intelligence approach for PM2.5 interpretation: A discovery of structural causal models

Abstract

Understanding the causal mechanisms underlying PM_2.5 generation is critical for developing effective prevention strategies, necessitating an approach that goes beyond prediction and seeks deeper causal explanations to support decision-making. This study addresses these concerns through a novel causal artificial intelligence framework employing structural causal models (SCMs) to interpret PM_2.5 dynamics. The research uncovers hidden cause-and-effect relationships between meteorological factors and PM_2.5 exposure by leveraging a data-driven causal structure discovery approach, effectively representing complex data-generating processes. The proposed SCMs undergo systematic validation across two critical dimensions: demonstrating human-like intelligence understanding and achieving significant alignment with real-world observations. The PC-based SCM is particularly outstanding when compared to other algorithms like GES- and Chow-Lui-based SCMs, delivering a remarkable performance in discovering cause-and-effect relationships with an F-measure of approximately 80 % compared to the gold-standard SCM. Statistical validation provided robust evidence of the model's reliability, with fit indices—including NFI, TLI, CFI, GFI, and AGFI—reaching approximately 0.98 and RMSEA approximating 0.05. These findings demonstrate that SCM can encode human-like reasoning and naturally align with real-world meteorological systems. This method is especially effective for urban air quality monitoring, where accessible meteorological data and transparent causal relationships are essential. Its capacity to inform evidence-based policy decisions makes it a powerful tool for creating intelligent decision-support systems in PM2.5 analysis and environmental mitigation strategies.