Developing an Explainable Variational Autoencoder (VAE) Framework for Accurate Representation of Local Circulation in Taiwan

• Published in: Journal of geophysical research. Atmospheres 2024-06, Vol.129 (12), p.n/a
• Main Authors: Hsieh, Min‐Ken, Wu, Chien‐Ming
• Format: Article
• Language: English
• Subjects: Accuracy; Circulation; Circulation patterns; Climate and weather; Climate change; Climate change scenarios; Climate prediction; Climatic conditions; deep generative model; deep learning; explainable artificial intelligence; Flow distribution; Flow pattern; Fluid flow; large eddy simulation; local circulation; Local flow; Neural networks; Representations; Simulation; Synoptic weather conditions; Training; Uncertainty; Vortices; Weather; Weather conditions; Weather patterns; Wind; Wind speed
• ISSN: 2169-897X; 2169-8996
• DOI: 10.1029/2024JD041167

This study develops an explainable variational autoencoder (VAE) framework to efficiently generate high‐fidelity local circulation patterns in Taiwan, ensuring an accurate representation of the physical relationship between generated local circulation and upstream synoptic flow regimes. Large ensemble semi‐realistic simulations were conducted using a high‐resolution (2 km) model, TaiwanVVM, where critical characteristics of various synoptic flow regimes were carefully selected to focus on the effects of local circulation variations. The VAE was constructed to capture essential representations of local circulation scenarios associated with the lee vortices by training on the ensemble data set. The VAE's latent space effectively captures the synoptic flow regimes as controlling factors, aligning with the physical understanding of Taiwan's local circulation dynamics. The critical transition of flow regimes under the influence of southeasterly synoptic flow regimes is also well represented in the VAE's latent space. This indicates that the VAE can learn the nonlinear characteristics of the multiscale interactions involving the lee vortex. The latent space within VAE can serve as a reduced‐order model for predicting local circulation using synoptic wind speed and direction. This explainable VAE binds the physical reasoning to the predictions of the local circulation that ensures the physical examination of the uncertainty in accelerating the local weather assessments under various climate change scenarios. Plain Language Summary This research introduces an advanced neural network framework for generating high‐fidelity local flow patterns in Taiwan. This framework, known as an explainable variational autoencoder, can accurately simulate how wind patterns of synoptic weather conditions interact in this region. We used detailed simulations to train the variational autoencoder, ensuring it captures the complex relationships between local flow and larger‐scale weather patterns. By training on the detailed simulations, the variational autoencoder learned and represented these large‐scale weather patterns in a way that helps maintain the physical relationship between local flow prediction and the large‐scale weather patterns. One of the key outcomes of this study is the development of a reduced‐order model. This simplified model takes advantage of what we have learned about complex weather interactions and can quickly predict local weather under different conditions.