Causal inference for quantifying chemical–dynamical pathways controlling tropical middle stratospheric ozone variability

Understanding the chemical–dynamical interactions controlling ozone (O 3 ) variability in the tropical middle stratosphere is essential for interpreting short-term trends and their sensitivity to dynamical fluctuations. This study applies a process-oriented causal inference framework that combines causal discovery and causal effect estimation. This approach integrates qualitative physical knowledge through a causal graph applied to satellite observations and a chemistry-transport model (CTM) simulation, using monthly data for the period 2004–2021. Causal inference robustly identifies a dominant chemical–dynamical pathway, in which variability in residual vertical velocity ( w * ) modulates nitrous oxide (N 2 O), subsequently affecting nitrogen dioxide (NO 2 ) and ultimately O 3 . Estimates of direct causal effect capture that O 3 variability is dominated by this indirect NO 2 -mediated pathway, while the direct influence of w * on O 3 is weak. The total causal effect (direct and mediated) peaks at a lag of approximately two-three months, indicating the cumulative impact of persistent, vertically coupled w * anomalies associated with the QBO. Regime-oriented analysis applied to the observations reveals that the chemical links (N 2 O–NO 2 and NO 2 –O 3 ) strengthen during westerly QBO shear compared to easterly shear. Our study highlights the pivotal role that causal inference can play in disentangling complex chemical-dynamical influences on O 3 , complementing traditional statistical methods. This approach lays the foundation for broader applications in stratospheric chemistry, where the understanding of various feedback pathways remains uncertain. By discovering and quantifying causal links, this methodology can be adapted to address open questions with environmental and societal relevance. Integrating causal reasoning into data-driven science can enhance process understanding and also strengthen the synergy between machine learning and statistical methods in Earth and environmental sciences.