Collaborative Research: Advancing Statistical Surrogates for Linking Multiple Computer Models with Disparate Data for Quantifying Uncertain Hazards

2014年导致43人死亡的华盛顿奥索山体滑坡，以及2010年导致欧洲航空停运的Eyjafjallajökull（冰岛）火山喷发的火山灰，都是罕见的灾难性地球物理事件的例子。它们的罕见性质使得这类事件几乎不可能预测，如果预测仅仅基于以前的观察。为了捕捉罕见事件，研究人员必须依靠复杂的物理和数学模型，而这些模型通常需要大量的计算资源来运行。此外，这类事件可能最好由一系列不同尺度上不同现象的不同模型来描述。例如，研究人员可能需要将降雨模型、边坡破坏模型和滑动碎片模型结合起来，以创建滑坡事件的总体模型。本研究的主要目标是开发有效的统计和计算策略来结合这些模型，从而推进灾害预测的最新技术。直接的基于模拟的危害评估需要成千上万个相互关联的时空模拟。此外，为了在危害评估中发挥最大作用，这些模拟应得到观测数据集的信息和验证，这些观测数据集本身可以从稀疏数据（罕见事件）到大量数据（例如卫星数据），并对新出现的情景进行探索。使问题复杂化的是，感兴趣的问题的许多特征要么特征不明确，要么不可预测，人们希望在每个特征的一系列值下运行模拟程序；这很快导致需要为数十万或数百万种不同的参数值和条件组合运行模拟程序（可能需要数小时才能完成）。根本没有足够的时间或足够的计算能力让这种蛮力方法取得成功。为了解决刚才描述的情况，pi将继续为大量时空模拟器数据开发并行部分模拟器，允许在地球物理模拟中常用的自适应时空网格上构建模拟器，为其输出创建平滑器，并允许使用减少的输入空间。pi将开始研究通过多个模拟器连接多个模拟器的策略。一种特别强大的连接模拟器的半分析方式将被追求，围绕该方法的准确性以及其在为并行部分模拟器设想的大数据场景中实现的可能性而产生的各种研究问题。pi还将开始研究提取（几乎）最优基集的技术、数据约简方法和算法方法，以加速模拟器的构建，所有这些都有助于更健壮地处理大型数据集。这些新方法将为快速构建基于概率的级联地球物理事件灾害预测图提供工具。快速预测图允许最终用户在各种任意场景下进行危害分析。此外，这种新方法将能够快速评估认知的不确定性。这种方法在基于科学的决策支持方面取得了巨大的进步。

The Oso, Washingon landslide of 2014, which resulted in 43 fatalities, and the ash plumes from the Eyjafjallajökull (Iceland) eruption of 2010, which shut down air travel in Europe, are examples of rare and catastrophic geophysical events. Their rare nature makes such events nearly impossible to forecast, if forecasts are based only on previous observations. To capture rare events, researchers must rely on complex physical and mathematical models that often require significant computational resources to exercise. Furthermore, events like these may be best described by a series of different models of different phenomena at different scales. For example, a researcher may need to combine a model of rainfall, a model of slope failure, and a model of sliding debris to create on overall model for a landslide event. The main objective of this research is the development of efficient statistical and computational strategies to combine such models, thus advancing the state of the art in hazard forecasting.Direct simulation-based hazard assessment would require thousands to tens of thousand of linked, space-time simulations. Furthermore, to be of most use in hazard assessment, these simulations should be informed and validated by observational data sets, which themselves can range from sparse data (rare events) to massive data (e.g. satellite data), and explored for emerging scenarios. To complicate the matter, a number of features of the problems of interest are either poorly characterized or unpredictable, and one would like to run the simulation programs at a range of values of each of them; this quickly leads to a perceived need to run a simulation program (which may take hours to complete) for hundreds of thousands or millions of different combinations of parameter values and conditions. There simply is not enough time or enough computing power for such a brute force approach to succeed. To tackle the situation just described, the PIs will continue to develop parallel partial emulators for massive space-time simulator data allowing emulator construction on the adaptive space-time grids commonly used in geophysical simulations, creating smoothers for their output, and enabling the use of reduced input spaces. The PIs will begin the investigation of a strategy for linking multiple simulators via multiple emulators. A particularly powerful semi-analytic way of linking emulators will be pursued, with a variety of research questions arising centering around the accuracy of the method, as well as the possibility of its implementation in the huge data scenario envisaged for the parallel partial emulator. The PIs will also begin to investigate techniques to extract (nearly) optimal basis sets, data reduction methods, and algorithmic approaches to accelerate the construction of emulators, all of which contribute to a more robust handling of large datasets. These new methodologies will provide tools to rapidly construct probability-based hazard forecast maps for cascading geophysical events. Rapid forecast maps allow end users to perform hazard analysis under a wide variety of aleatoric scenarios. Furthermore this new methodology will enable fast assessment of epistemic uncertainties. This approach constitutes a dramatic improvement in scientifically-based decision support.