Abstract Liquefaction is a major cause of coseismic damages, occurring irregularly over hundreds or thousands of square kilometers in large earthquakes. Large variations in the extent and location of liquefaction have been observed in recent earthquakes, motivating the need for prediction methods that consider the spatial heterogeneity of geologic deposits at a regional scale. Contemporary regional-scale liquefaction hazard analyses are typically performed using only surficial data, which does not address the complicated subsurface mechanics and spatial variability associated with artificial fill and natural soil deposits. In this study, we develop a probabilistic, regional-scale, subsurface model using data from hundreds of borings to better understand subsurface conditions that could influence liquefaction. We then use this subsurface sample database to train Gaussian process models, yielding 3D independent random fields of groundwater depth, soil plasticity, and penetration resistance for each geologic unit. We incorporate the Gaussian process models into probabilistic liquefaction triggering procedures, producing 3D estimates of the probability of liquefaction for an example study area in Portland, Oregon. Near sampling locations, the variance of the Gaussian process models approaches the variance of site-specific liquefaction triggering procedures. Conversely, when no sample data are nearby to condition a Gaussian process, the variance approaches the marginal variance of the entire recorded dataset. Thus, the procedure described in this study unifies probabilistic site-specific and regional-scale liquefaction triggering procedures and provides an important step towards quantitative liquefaction hazard assessments for regionally distributed infrastructures, such as levees, pipelines, roadways, and electrical transmission facilities.

中文翻译：

---

使用 3D 高斯过程的概率区域尺度液化触发建模

摘要 液化是造成同震破坏的主要原因，在大地震中不规则地发生在数百或数千平方公里的范围内。在最近的地震中观察到液化程度和位置的巨大变化，促使需要考虑区域尺度地质沉积物空间异质性的预测方法。当代区域尺度的液化危害分析通常仅使用地表数据进行，这并没有解决与人工填充物和天然土壤沉积物相关的复杂地下力学和空间变异性。在这项研究中，我们使用来自数百个钻孔的数据开发了一个概率、区域尺度的地下模型，以更好地了解可能影响液化的地下条件。然后我们使用这个地下样本数据库来训练高斯过程模型，为每个地质单元生成地下水深度、土壤可塑性和渗透阻力的 3D 独立随机场。我们将高斯过程模型纳入概率液化触发程序，为俄勒冈州波特兰的一个示例研究区域生成液化概率的 3D 估计。在采样位置附近，高斯过程模型的方差接近特定地点液化触发程序的方差。相反，当附近没有样本数据来调节高斯过程时，方差接近整个记录数据集的边际方差。因此，