應用可解釋深度學習識別大氣過程中的非線性形態特徵

Abstract

近年來，深度學習技術在大氣科學領域研究中的應用迅速增長，然而，這些模型往往被視為缺乏物理透明度的「黑盒子」，限制了其在科學推論上的可靠性。本論文旨在建立一個可解釋的深度學習框架，提出以「形態學（Morphology）」作為連結數據驅動預測與大氣物理過程的關鍵指標。透過捕捉物理變數在空間上的非線性形態特徵，我們得以在不同尺度下揭示大氣動力過程的演變機制。 本研究針對三個不同的大氣尺度，分別採用了三種深度學習架構來驗證此框架的可行性。首先，在大尺度方面，針對突發性平流層增溫（SSW）現象，我們利用「卷積變分自動編碼器（Convolutional Variational Autoencoder）」構建了平流層極地渦旋的潛在變數空間相位圖。結果顯示，VAE 能有效捕捉渦旋從位移型到分裂型的非線性連續演變路徑，其表現優於傳統的線性主成分分析（PCA），為極地渦旋的形態演變研究，以及未來的預測研究提供了新的視角。 其次，在中尺度方面，針對熱帶對流聚合現象，我們開發了一套迭代式特徵移除的卷積神經網路（CNN）框架。透過反覆訓練與遮蔽輸入資料中的顯著特徵（如平均雲水含量），我們嘗試找出隱藏在神經網路決策背後的關鍵形態因子。研究發現，除了雲覆蓋率外，雲的邊緣複雜度（碎形維度）是模型辨識對流是否集結的一個重要的非線性特徵。 最後，在小尺度方面，針對深對流中的冷池（Cold Pool）動力過程，我們應用U-Net模型，嘗試解決全球風暴解析模式（GSRMs）中對於冷池的解析度不足的問題。本研究利用高解析度大渦模擬（LES）資料，成功從低解析度的環境場中重建出高解析度的冷池形態與強度分佈。同時，敏感度分析進一步顯示近地面的動力場對於冷池邊界的重建有一定的重要性，其影響力在此尺度下的冷池結構重建上大於熱力變數。 綜合以上，本論文展示了深度學習模型在經過適當設計後，不僅能作為預測工具，更能成為探索大氣非線性形態特徵的分析工具。此框架透過捕捉大尺度物理場演變過程、擷取中尺度對流特徵、到嘗試重建次網格細節，期望能為提升數值模式的參數化的未來發展與深度學習可解釋性提供貢獻。

The application of deep learning in atmospheric sciences has expanded rapidly in recent years. However, these models are often treated as "black boxes" lacking physical transparency, which limits their reliability for scientific inference. This dissertation aims to establish an explainable deep learning framework by proposing "morphology" as a key physical indicator to bridge the gap between data-driven predictions and atmospheric physical processes.

This research validates the proposed framework through three distinct atmospheric scales using appropriate deep learning architectures. First, at the large scale, focusing on Sudden Stratospheric Warmings (SSWs), we use a Convolutional Variational Autoencoder (VAE) to construct a latent space phase diagram of the stratospheric polar vortex. The results demonstrate that the VAE effectively captures the nonlinear continuous evolution of the vortex—transitioning between displacement and splitting events—outperforming Principal Component Analysis (PCA) and providing a novel perspective on vortex dynamics.

Second, at the mesoscale, investigating tropical convective aggregation, we develop an iterative feature-removal framework using Convolutional Neural Networks (CNNs). By iteratively training the model and masking dominant features (such as average cloud water path) from the input data, we aim to identify and extract the hidden morphological factors driving the network's decisions. This study reveals that, beyond cloud coverage, the complexity of cloud edges (fractal dimension) serves as an important nonlinear feature for identifying the occurrence of the convective aggregation. Finally, at the small scale, addressing cold pool dynamics within deep convection, we apply a U-Net model to deal with the resolution gap in Global Storm-Resolving Models (GSRMs). Using high-resolution Large-Eddy Simulation (LES) data, we successfully reconstruct high-resolution cold pool morphology and intensity distributions from coarse-grained environmental fields. Sensitivity analysis further reveals that the near-surface dynamic field (wind divergence signals) is important for reconstructing cold pool boundaries, playing a more critical role than thermodynamic variables alone.

In conclusion, this dissertation demonstrates that when properly designed, deep learning models can serve not only as prediction tools but also as powerful physical analysis instruments for exploring nonlinear atmospheric morphology. By representing large-scale evolution, extracting mesoscale features, and reconstructing subgrid details, this framework is expected to provide a foundation for improving parameterization schemes and enhancing physical interpretability of deep learning research in atmospheric sciences in the future.