發展可解釋機器學習框架: 臺灣背風渦漩局地環流之應用

Abstract

本研究旨在透過結合高解析度物理模型TaiwanVVM和變分自編碼器（VAE）神經網絡，在機器學習框架的預測能力和可解釋性之間取得平衡。我們聚焦於東亞冬季季風期間，挑選了此期間可影響臺灣局地環流的各種綜觀尺度環境，使用解析度為2公里的TaiwanVVM進行系集模擬，以研究綜觀尺度條件對於臺灣局地環流的影響；利用這組模擬結果作為VAE之訓練資料集，使其得以捕捉背風渦旋主導的局地環流型態。VAE的特徵空間(latent space)有效地涵蓋了控制局地環流的綜觀條件，這與流體力學上臺灣局地背風渦旋形成原因的物理解釋相符合。這代表VAE能夠學習與背風渦旋形成相關的多重尺度非線性交互作用的現象。當VAE的特徵空間可以解釋為控制各種局地環流情境的綜觀環境流場時，此結果即建構了利用綜觀環境風速和風向來預測局地環流的降維模型。該模型能有效預測由大尺度環流驅動的局地環流，並可應用於未來暖化情境下評估臺灣局地環流與伴隨的空污天氣。本研究所提出的框架提供了一種具有物理解釋性並有效率的方法來評估全球暖化在局地造成的影響，有助於氣候適應政策的制定。本研究亦展示了透過機器學習可在高解析度模擬結果中學習其物理關聯，從而促進機器學習應用於大氣科學的可解釋性和可靠性。

This study aims to balance predictive power with interpretability by integrating the high-resolution physical model TaiwanVVM with the variational autoencoder (VAE) neural network framework. Focusing on local circulation in Taiwan during the East Asian Winter Monsoon (EAWM), we conduct large ensemble semi-realistic simulations using TaiwanVVM at a high resolution of 2 km, selecting critical characteristics of various synoptic flow regimes to examine their effects on local circulation. The VAE was trained on this ensemble dataset to capture essential representations of local circulation scenarios associated with lee vortices. The VAE's latent space effectively encapsulates synoptic flow regimes as controlling factors, aligning with the physical understanding of Taiwan’s local circulation dynamics. This indicates that the VAE can learn the nonlinear characteristics of multiscale interactions involving the lee vortex. As the latent space of VAE is interpreted as the physical parameters of upstream flow regimes driving the variability of local flows in large ensemble simulations, it serves as a reduced-order model for predicting local circulation based on synoptic wind speed and direction. This model efficiently predicts local circulation patterns driven by synoptic flow regimes and is applicable to future climate projections, enabling the assessment of global warming's impact on local circulation and associated future pollution weather. The proposed framework provides a physically explainable and efficient method for aiding policy making in climate change adaptations. This study demonstrates that our machine learning framework can learn physical connections from high-resolution ensemble simulations, thereby promoting interpretability and reliability in their application.