Statistical Methods for Ice Sheet Projections using Large Non-Gaussian Space-Time Data Sets and Complex Computer Models

该项目的目标是通过使用新的统计方法来研究南极冰盖的过去、现在和未来的行为，将最先进的冰盖模型和观测数据结合起来。诸如PSU 3D冰盖模型之类的冰盖模型可以用来理解驱动南极冰盖行为的过程之间的相互作用。它们还可以用来预测冰盖未来的行为，这对预测气候变化，特别是海平面上升非常有意义。了解未来海平面的上升不仅对科学家，而且对政策制定者都很重要。这项跨学科的研究项目结合了统计方法、冰盖模拟、计算和强迫情景方面的专业知识。由于冰盖模型输出是高维、非高斯和时空的，现有的计算机模型模拟和校准方法要么不适用，要么在计算上不可行。构建计算上方便的方法是具有挑战性的，这些方法足够灵活，可以利用大型数据集中的所有信息，并考虑到不确定性、复杂的错误和相关性。我们将开发新的统计模型和算法来应对这一挑战。我们的统计方法将建立在马尔可夫随机场模型(高斯和非高斯)、复合似然和相关似然近似方法以及推广到指数族的主成分方法的基础上。目标有两个：(1)开发新的统计和计算方法，用于使用高维非高斯时空数据对复杂计算机模型进行仿真(随机逼近)和校准(输入参数学习)；(2)了解驱动南极冰盖行为的主要科学过程，并首次在这方面利用过去20 000年来地质数据的信息对该区域未来的冰消退作出预测。我们将把重点放在南极洲西部的阿蒙森海部分，那里有迅速后退和变薄的松岛和斯韦茨冰川；这一部分目前是南极海平面上升的最大贡献者，被认为特别容易受到未来急剧消退的影响。将我们的冰盖模型与这一领域在过去20,000年的去冰川衰退期间的地质数据进行严格的统计比较，将产生对未来进一步消退的更有力和更有信心的预测。我们还将探索南极冰盖对大气-海洋过程的敏感性，这是由耦合地球系统模型估计的情景驱动的。驱动区域冰盖的过程将在共同体地球系统模型(CESM)的组成部分内进行调查。冰盖融化会影响海平面上升，并可能对人类和生态系统产生重大影响。因此，对科学家和政策制定者来说，理解它们是至关重要的。这里开发的统计方法和计算工具将广泛适用于具有高维输出的计算机模型常见的一系列学科；这些领域包括生态学、水文学、机械工程和天文学。我们的高效方法和软件将允许我们(和其他人)常规地将更复杂的模型适合于比目前可行的更大的数据集，从而在得出科学结论时使用尽可能多的信息。这将减少不确定性，从而将数据集的大小转化为资产。

The goal of this project is to study the past, current, and future behavior of the Antarctic ice sheet by using novel statistical methods for combining information from both state-of-the-art ice sheet models and observational data. Ice sheet models such as the PSU 3D ice sheet model may be used to understand the interplay between the processes that drive the behavior of the Antarctic ice sheet. They can also be used to project the future behavior of the ice sheet, which is of great interest in projections of climate change, particularly sea level rise. Learning about future sea-level rise is important not only to scientists but also policymakers. This interdisciplinary research project combines expertise in statistical methodology, ice sheet modeling, computing, and forcing scenarios.Because ice sheet model outputs are high-dimensional, non-Gaussian, and spatio-temporal, existing approaches for computer model emulation and calibration either do not apply or are computationally infeasible. It is challenging to build computationally expedient methods that are flexible enough to utilize all the information in large data sets and account for uncertainties, complicated errors, and dependencies. We will develop new statistical models and algorithms to address this challenge. Our statistical methodology will build upon Markov random field models (Gaussian and non-Gaussian), composite likelihood and related likelihood approximation methods, and principal components methods generalized to exponential families. The objectives are twofold: (i) developing new statistical and computational methods for emulation (stochastic approximation) of and calibration (input parameter learning) for complex computer models with high-dimensional non-Gaussian space-time data, and (ii) learning the major scientific processes driving the behavior of the Antarctic ice sheet and making projections for ice retreat in this region in the future using, for the first time in this context, information from geologic data over the past 20,000 years. We will focus on the Amundsen Sea Sector of West Antarctica, which contains the rapidly retreating and thinning Pine Island and Thwaites Glaciers; this sector is currently the largest Antarctic contributor to sea-level rise and is considered particularly vulnerable to drastic future retreat. Rigorous statistical comparisons of our ice-sheet model versus geologic data in this sector during the deglacial recession of the last 20,000 years will yield more robust and confident predictions of further future retreat. We will also explore sensitivity of Antarctic ice sheet response to atmosphere-ocean processes as driven by scenarios estimated from coupled earth system models. The processes driving the regional ice sheet will be investigated within the components of the Community Earth System Model (CESM). Ice sheet melting influences sea level rise and can have major impacts on human and ecological systems. Hence, understanding them is paramount to scientists and policy makers. The statistical methodology and computational tools developed here will be widely applicable across a range of disciplines where computer models with high-dimensional output are common; these fields include ecology, hydrology, mechanical engineering, and astronomy. Our efficient methods and software will allow us (and others) to routinely fit more sophisticated models to larger data sets than currently feasible, thus using as much information as possible when drawing scientific conclusions. This will allow for reduced uncertainties, thereby turning the size of the data sets into an asset.