[A-AS03] Extreme Events and Mesoscale Weather: Observations and Modeling

[AAS03-21] A physically interpretable AI framework (AI-TaiwanVVM) for predicting future pollution weather in Taiwan

*Min-Ken Hsieh¹, Chien-Ming Wu¹ (1.Department of Atmospheric Sciences, National Taiwan University)

Keywords:explainable AI, flow over complex topography, pollution weather, storyline approach

Fine particulate matter (PM_2.5; particles with a diameter of 2.5 micrometers or less) pollution is a serious environmental issue as it is a leading cause of disease and premature death worldwide. Considering the pollutant transport scenarios over a limited area with complex topography such as Taiwan, local circulation can significantly influence the pollutant transport pathway. It is necessary to evaluate the details of the local circulation in Taiwan to better assess the potential severe pollution weather scenarios in Taiwan under the warming climate. However, the traditional numerical downscaling approaches, which provide high-resolution details of the meteorological conditions within a focal area, usually rely on massive computational resources. We present "AI-TaiwanVVM," an innovative framework integrating variational autoencoders (VAEs) with semi-realistic TaiwanVVM simulations to predict local circulation in Taiwan while preserving physical interpretability. VAEs provide a structured and continuous latent space for extracting low-dimensional representations from high-dimensional meteorological data, enabling better physical interpretability in developing Artificial Intelligence (AI) tools for weather prediction. Combining VAEs with TaiwanVVM simulations, the framework captures variability in local circulation associated with lee vortices under diverse synoptic conditions. This integration aligns with the scenario-based storyline approach to reduce the uncertainty in utilizing traditional downscaling methods to evaluate the local weather change under global warming. AI-TaiwanVVM ensures that the generated scenarios align with physical principles, enabling the construction of reduced-order models that efficiently predict details of local pollution weather variability. The framework's emphasis on physical interpretability is critical for developing AI models and advancing our understanding of local pollution weather and their response to changing synoptic conditions in a warming climate.