應用機器學習於土壤液化潛能分類之預測

Abstract

臺灣位處環太平洋地震帶，頻繁的地震經常對不論是建築物或是民生經濟都帶來相當程度的威脅，地震帶來的循環荷載使土壤顆粒間的孔隙水壓上升，土層失去原有的有效應力，造成土壤液化的現象，為近十年來火熱的議題。 本研究為探究不同類型地震作用於土層之影響，蒐集了共743筆地震訊號，包含淺殼層型、海溝型、近斷層脈衝型及非近斷層脈衝型等地震，土層部分則分為數值土層及臺灣現地土層兩類，模擬土層部分主要由砂土構成，依相對密度、地層深度、土壤分層數量等差異一共生成100種土層；現地土層則是透過臺灣中南部的真實土層資料建立，共選取7個土層，以上述資料用OpenSees建立數值模型，進行了79501組三維有效應力地盤反應分析，討論不同地震、土層下對超額孔隙水壓比 (r_u) 及最大剪應變 (γ_max) 的影響。 將地盤反應分析的結果整合，分別建立數值土層及臺灣現地土層的資料庫，內容包括土壤參數如深度、相對密度；各地震之地震動強度參數 (ground motion intensity measures, IMs) 如PGA、PGV、CAV等參數以及r_u和γ_max，使用機器學習方法：可解釋提升機器 (explainable boosting machine, EBM)、隨機森林 (random forest, RF)、極限梯度提升 (extreme gradient boosting , XGBoost) 作訓練，以土壤參數及單一IM作為輸入特徵，r_u、γ_max及液化種類為預測目標，比較不同演算法和資料庫訓練出的差異，並將訓練出的模型套用至歷史液化案例上，觀察各自的預測表現，其中在不論r_u或γ_max的預測上，以PGV作為輸入特徵參數的模型皆表現出較高的預測能力，顯現出PGV與兩者間的高度相關。

Taiwan is located along the Pacific Ring of Fire, where frequent earthquakes pose significant threats to both infrastructure and the general economy. The cyclic loading induced by ground motions lead to an increase in pore water pressure between soil particles, causing the soil to lose its original bearing capacity and resulting in soil liquefaction—a topic that has garnered much attention in the past decade. This study aims to investigate the effects of different types of earthquakes on soil layers. A total of 743 ground motion records were collected, including shallow crustal, subduction zone, near-fault pulse-like, and non-pulse-like earthquakes. Two categories of soil layers were considered: synthetic and field. The synthetic soil layers, primarily composed of sandy soils, were generated based on variations in relative density, depth, and number of soil strata, resulting in 100 different configurations. The field soil profiles were established using actual geotechnical data from central and southern Taiwan, consisting of 7 selected sites. Using the above data, numerical models were constructed in OpenSees to perform 79,501 three-dimensional effective stress site response analyses. The impact of various ground motions and soil conditions on excess pore water pressure ratio (r_u) and maximum shear strain (γ_max) was analyzed. The results of the site response analyses were compiled into two separate databases for synthetic and field soil layers. These databases included soil parameters (e.g., depth, relative density), ground motion intensity measures (IMs) such as PGA, PGV, and CAV, as well as the response metrics r_u and γ_max. Three machine learning models—Explainable Boosting Machine (EBM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—were trained using soil parameters and a single IM as input features, with r_u, γ_max, and liquefaction classification as prediction targets. The performance of each model was compared based on different datasets and algorithms. Furthermore, the trained models were applied to historical liquefaction cases to evaluate their predictive capability. Among all IMs considered, models using PGV as the input feature consistently demonstrated superior prediction performance for both r_u and γ_max, indicating a strong correlation between PGV and these two response parameters.