Using a Physical Interpretable <scp>AI</scp> Framework to Predict Local Pollution Weather Deterioration in Taiwan Under <scp>TaiESM1 SSP585</scp> Projection

ABSTRACT This study leverages an explainable artificial intelligence (AI) model trained on high‐resolution physical simulations over complex terrain of Taiwan to efficiently predict and interpret local pollution deterioration scenarios under TaiESM1 SSP585 climate projection. As the lee vortex in Taiwan controls the variability of the local pollution deterioration scenarios, TaiESM1 simulations under the SSP585 scenario suggest notable increases in lee‐vortex days (LVDs) around Taiwan, rising from 597 days (2021–2030) to 689 days (2091–2100). These synoptic conditions, characterised by weakened north‐easterly and enhanced south‐easterly winds, significantly modulate local pollutant transport pathways. This AI model, named AI‐TaiwanVVM, effectively reconstructs kilometre scale circulation patterns based on upstream flow regimes, revealing substantial increases in NO x accumulation, particularly in northern Taiwan and the Greater Taipei Area. The observed pollution deterioration can be physically interpreted as a consequence of the intensifying influence of the Pacific subtropical high during the spring transition. Furthermore, AI‐driven daily scenario analyses demonstrate how shifts towards more frequent and prolonged weak‐wind conditions exacerbate pollutant retention. A representative five‐day scenario from April 2092 is discussed to illustrate AI‐TaiwanVVM's capacity to simulate the critical role persistent synoptic conditions play in escalating local air pollution severity. These results underscore AI‐TaiwanVVM's effectiveness in rapidly predicting local air quality risks, highlighting its potential as a valuable tool for guiding adaptive air quality management under future climate conditions.