| 摘要: | 地震預警系統已經被廣泛地使用以減少傷亡以及經濟財產損失。中央氣象局目前利用區域型地震預警系統，對台灣發布地震預警資訊，島內地震預警時間為10秒，而島外地震為18秒。為了提高系統效率，大數據機器學習結合台灣密集的觀測資料有機會縮短預警時間，直接預測震度分布而無需計算地震事件的規模和位置，並減少盲區的範圍。然而，台灣地區中大型地震(M6+)資料的稀缺，將會使得機器學習模型訓練時產生偏差導致不穩定的預估。為了填補波形之不足，本研究將評估利用經驗格林函數(Empirical Green’s Function, EGF)法以及應變格林張量(Strain Green’s Tensor, SGT)法，用於模擬台灣中大型地震之波形的可行性。在EGF方法中，我們嘗試將小規模地震(M3~5+)波形，利用地震學原理放大為大規模地震(M6+)波形。此方法中，地震波之路徑項及場址項均包含於小規模地震波形。另一方面，基於物理的SGT方法可以對於台灣地區的特徵地震進行情境模擬，包含設定特定破裂尺寸、破裂速度、破裂方向、震源機制等震源參數，以及考慮三維速度構造模型與場址效應，適用於模擬規模更大之地震(M6~7+)波形。本文選擇了台灣三個M6+地震事件，即2013年南投、2016年美濃和2019年花蓮地震，驗證這兩種方法。EGF結果指出，在0.2~1 Hz頻段下，花蓮及南投地震的模擬震度可以與觀測震度良好符合，美濃地震的結果則顯示較高的震度殘差，這是由於EGF地震之震源沒有像美濃地震那樣強烈的破裂方向性，另外，可模擬合成波形之測站範圍會受限於記錄EGF地震波形之測站。而SGT結果顯示合成波形之振幅顯著受到速度模型偏差與淺層場址效應之影響，透過這兩項因素之修正，花蓮及南投地震的模擬震度可以在0.2~1 Hz頻段良好符合觀測震度，美濃地震則顯示真實的觀測震度分布較模擬震度更為複雜。本研究結果表明，EGF法與SGT法皆可以提供M6+合成地震波形，以填補台灣地震資料庫的不足，並增加機器學習之訓練資料，對於傳統地震動預估式之迴歸分析、地震工程等相關研究也有所助益。;Earthquake early warning systems have been used to mitigate injuries and damage worldwide for many years. In Taiwan, the Central Weather Bureau (CWB) has operated a regional type system to issue warnings to public. The inland and offshore warning times are about 10 and 18 seconds, respectively. In order to improve the efficiency of the system, more and more studies account for that machine learning approaches could predict the intensity distribution without calculating the magnitude and location of the event. However, lack of seismic records for moderate-large (M6+) earthquakes may give unstable extrapolations while predicting intensities. To fill the lack, we demonstrate two approaches, the empirical Green′s function (EGF) and the strain Green′s tensor (SGT), in seismic waveform simulations for moderate-large earthquake scenarios in Taiwan. In the EGF approach, the waveforms of scenario M6+ events are simulated by the observed waveforms of M3~5+ events, which contain the path and site effects. On the other hand, the SGT approach allows us to perform synthetics of larger event (M6~7+) scenarios physically, considering potential source mechanisms (e.g., dimension, rupture speed, directivity, focal mechanism) and structures (e.g., 3-D velocity structure and site response). Here, three M6+ events in Taiwan, the 2013 Nantou, the 2016 Meinong, and the 2019 Hualien earthquakes, are selected to validate these two approaches. In the EGF approach, the intensities from the synthetic waveforms are similar to those from the observations in the frequency range of 0.2~1 Hz for the Hualien and the Nantou events. However, higher intensity residuals appear for the Meinong earthquake. It may be due to that source of the EGF events does not have such strong directivity as the Meinong earthquake. Besides, the spatial distribution of simulations will be restricted by the number of on-site observations of EGF events. As for the SGT approach, after correcting the velocity model bias and site effect, the synthetic waveforms could show good intensity simulations for the Hualien and the Nantou events in the same frequency range, while the actual distribution of strong motion for the Meinong event is more complex than the simulation result. Overall, our results suggest that both approaches could fill the lack of seismic records for moderate-large earthquakes. These synthetic data have the potential to improve the machine learning in early warning systems, and can further be utilized to refine the existing ground motion prediction equation as well as applied to earthquake engineering research. |
