Collaborative Research: Fusing Massive Disparate Data and Fast Surrogate Models for Probabilistic Quantification of Uncertain Hazards

减轻火山喷发、地震或传染病等自然灾害的影响，取决于我们在灾害风险发生前准确量化风险的能力。该项目将解决这一挑战，并开发一种新的计算上可行的框架，以整合不同的现场观测和计算机模拟。新的框架将大幅提升自然灾害量化的计算效率。其中一个试验田将是2018年夏威夷基拉韦厄火山喷发，那次喷发造成23人受伤，700多所房屋被毁。在这次活动中，将使用来自不同来源的广泛的现场观测，如雷达卫星、全球导航卫星系统接收器、钻孔倾斜仪和地震仪，以及大规模的计算机模拟，来分析火山危险量化的方法。该项目中开发的方法将在开放源码软件中实施，可供广大科学家和工程师使用。该项目还包括对研究生和本科生的培训。精确量化不确定的自然灾害的第一个主要障碍是计算机模拟的计算可扩展性，因为它们通常需要在具有多维输入的海量时空区域上数值求解偏微分方程组。这一挑战将通过开发高斯过程(GP)仿真器作为计算上可行的替代模型来克服，以近似计算机实验的结果。这种方法很有吸引力，因为它不仅包括相对于坐标数量具有线性计算顺序的并行预测，而且它还利用坐标之间的相关性来实现快速预测采样。第二个计算挑战是融合来自多个来源的不同数据，以校准物理模型。该项目将通过量化数据处理中的不确定性和估计物理模型与现实之间的差异来解决这一挑战，以便进行数据整合。虽然这个项目专注于自然灾害量化方面的应用，但新的GP仿真器、用于模型校准的计算工具和数据集成方法将更广泛地扩展数据科学和机器学习算法的适用性。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Mitigating the impact of natural hazards, such as volcanic eruptions, earthquakes, or infectious diseases, rests on our ability to accurately quantify hazard risks in advance of their occurrence. This project will tackle this challenge and develop a new computationally feasible framework to integrate disparate field observations and computer simulations. The new framework will deliver substantial upgrades in computational efficiency for natural hazard quantification. One testbed will be the 2018 eruption of the Kilauea Volcano in Hawaii, which injured 23 people and destroyed more than 700 dwellings. For this event, extensive field observations from disparate sources, such as radar satellites, global navigation satellite system receivers, borehole tiltmeters, and seismometers, as well as large-scale computer simulations, will be used to analyze methods for volcanic hazard quantification. The methods developed in the project will be implemented in open-source software available to a wide community of scientists and engineers. The project is complemented by training for both graduate and undergraduate students. The first major roadblock for precisely quantifying uncertain natural hazards is the computational scalability of computer simulations, as they often require the numerical solution of partial differential equations on massive spatio-temporal domains with multi-dimensional input. This challenge will be overcome by developing Gaussian process (GP) emulators as a computationally feasible surrogate model to approximate outcomes of computer experiments. This approach is appealing because it not only includes parallel predictions with linear computational order with respect to the number of coordinates, but it also leverages the correlation between coordinates to enable fast predictive sampling. The second computational challenge is in fusing disparate data from multiple sources to calibrate physical models. The project will address this challenge by quantifying uncertainty in data processing and estimating the discrepancy between the physical model and reality to allow for data integration. While this project focuses on applications in natural hazard quantification, the new GP emulator, computational tools for model calibration, and data integration methods will more generally extend the applicability of data science and machine learning algorithms.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.