臺灣位於西北太平洋颱風主要侵襲路徑上，每年夏秋季經常受到颱風影響，伴隨強風、豪雨及風暴潮等災害。其中，風暴潮所造成之沿海淹水、港灣設施損壞及海岸侵蝕等問題，對沿岸地區之生命財產安全與經濟活動造成嚴重衝擊。因此，如何提升風暴潮預報之準確性與效率，已成為防災與海岸工程領域的重要研究課題。 風暴潮模擬結果高度依賴海面風場與氣壓場之準確性。然而，目前常用之參數化風場模型多建立於理想化與軸對稱假設，難以描述颱風通過臺灣時受中央山脈地形阻擋、繞流及通道效應影響所產生之風場變形、強度削弱與局部增強等現象；另一方面，高解析度數值天氣模式雖可提供受地形影響後之氣象場資訊，但其計算成本高且不利於即時預報應用。因此，發展兼具計算效率與物理合理性之颱風氣象場重建方法，具有重要之研究與實務價值。 本研究以ERA5再分析資料為基礎，建立臺灣周邊颱風氣象場資料庫，並利用颱風路徑、中心氣壓及最大風速等特徵作為模型輸入，學習臺灣地形影響下之颱風氣象場空間分布特徵。為評估不同機器學習方法於颱風氣象場重建之適用性，本研究首先建構卷積神經網路（Convolutional Neural Network， CNN）作為純資料驅動基準模型，並進一步發展物理資訊神經網路（Physics-Informed Neural Networks， PINNs）模型，進行颱風氣壓場以及風場重建。PINNs模型於訓練過程中結合觀測資料與物理約束條件，使模型在學習氣象場空間分布特徵之同時，亦能維持物理一致性，進而提升模型於極端條件下之穩定性與泛化能力。 研究結果顯示，相較於純資料驅動模型，PINNs能較準確重建颱風中心附近之強梯度氣壓場與風場結構，並有效捕捉受臺灣地形影響所形成之非對稱分布特徵。在氣壓場、風場及暴潮空間分布等評估指標上，PINNs於各項評估指標中均呈現較低之誤差與較高之空間分布一致性。進一步將重建之氣象場輸入COMCOT風暴潮模式進行模擬，結果顯示PINNs所重建之氣壓場及風場能夠合理再現暴潮空間分布特徵，且與ERA5驅動之模擬結果呈現相近之空間分布與變化特徵。 本研究證實，透過PINNs結合地形鎖定效應與物理約束之建模方法，可有效提升颱風風壓場重建能力，並具備快速重建氣象場之能力，可有效降低高解析度數值天氣模式所需之計算成本。未來可進一步精進，輔助於即時風暴潮預報系統，提供臺灣沿岸防災決策與風險管理之重要參考。;Taiwan is frequently affected by typhoons originating in the western North Pacific， which often bring strong winds， heavy rainfall， and storm surges. Among these hazards， storm surges can cause coastal flooding， damage to harbor facilities， and coastal erosion， posing serious threats to lives， property， and economic activities in coastal regions. Therefore， improving the accuracy and efficiency of storm surge forecasting is an important issue in disaster prevention and coastal engineering. The reliability of storm surge simulations strongly depends on the accuracy of surface wind and atmospheric pressure fields used as forcing inputs. However， conventional parametric typhoon wind field models are usually based on idealized and axisymmetric assumptions， making it difficult to represent the deformation， weakening， and localized enhancement of wind fields caused by the blocking， channeling， and flow-deflection effects of Taiwan’s Central Mountain Range. Although high-resolution numerical weather prediction models provide more realistic wind and pressure fields， their high computational cost limits their application in real-time forecasting. In this study， a typhoon wind and pressure field database was established using ERA5 reanalysis data over the Taiwan region. Typhoon track information， central pressure， and maximum wind speed were used as input features to reconstruct wind and pressure fields affected by Taiwan’s complex terrain. A Convolutional Neural Network (CNN) was first developed as a purely data-driven baseline model， followed by a Physics-Informed Neural Network (PINN) for typhoon wind and pressure field reconstruction. By incorporating physical constraints into the training process， the PINNs model improves physical consistency， model stability， and generalization capability under extreme typhoon conditions. The results show that the PINNs model outperforms the conventional CNN model in reconstructing strong pressure gradients and wind structures near the typhoon center， while also capturing asymmetric spatial patterns induced by Taiwan’s terrain. In terms of pressure field reconstruction， wind field reconstruction， and spatial structural similarity， the PINNs model achieves lower prediction errors and higher similarity. Furthermore， the reconstructed wind and pressure fields were used as forcing inputs for the COMCOT storm surge model. The resulting storm surge simulations reasonably reproduce the spatial distribution characteristics of storm surges and exhibit similar spatial distribution with simulations driven by ERA5 reanalysis data. Overall， this study demonstrates that integrating topography locking effects and physical constraints through PINNs can effectively improve the reconstruction of typhoon wind and pressure fields. The proposed framework enables rapid generation of physically consistent forcing fields and reduces the computational cost associated with high-resolution numerical weather prediction models. Therefore， it has strong potential for application in real-time storm surge forecasting and can support coastal disaster prevention， emergency response， and risk management in Taiwan.
