Type 1: Collaborative Research: Bayesian Hierarchical Climate Prediction LO2170174

柏林，1049064维克尔，1049093该项目的重点是应用贝叶斯推理开发气候和气候影响预测框架。贝叶斯分析是一种理论上有基础的技术，用于组合信息，从数据中学习，然后在表示这些步骤中的不确定性的同时形成预测。现代分层贝叶斯建模已被证明在复杂环境下非常有效。该策略涉及各种观测数据源的概率分布的规范，称为数据模型。此外，对于在不同时空尺度上相互作用的未知过程以及未知参数，建立了称为先验过程模型的概率分布。通过贝叶斯定理将过程的先验分布与数据模型相结合，得到未知量的后验概率分布。这一分布是预测的基础。对概率的依赖意味着不确定性管理是该方法的内置内容。我们方法的关键是将全球和区域尺度的气候模型输出纳入其中，就像它们是观测一样。然后将这些数据与相对简单但基于物理的先前模型相结合。与大规模气候系统模型相比，预测和不确定性量化所需的相应计算要容易几个数量级。概率论提供了使用分析来产生影响的预测和相关不确定性的技术。这些结果将作为决策支持的基础。当前气候研究中的一个关键挑战是在十年时间尺度上预测区域气候行为。对十年尺度气候的可靠预测是决策者选择适应战略以应对气候变化影响的决策支持的基石。关于气候的大量信息由气候系统模型产生。然而，模型结果受到各种不确定因素的影响。此外，这些模型在计算上要求非常高，并产生大量数据集，因此我们量化与基于模型的预测相关的不确定性的能力受到严重限制。为了应对这一挑战，我们开发了用于预测分析的贝叶斯技术，该技术以有效的方式使用气候模型输出，并处理不确定性。这种大规模气候系统模型、区域气候模型和简单气候模型的组合使用创造了气候科学中的协同环境。其次，由于我们的战略依赖于相对简单的计算，我们提供了产生特定影响和特定于当地或区域的预测信息的技术，以响应具有不同职责的各种决策者的需求。

Berliner, 1049064Wikle, 1049093 The focus of this project is the development of a climate and climate-impact prediction framework by applying Bayesian reasoning. Bayesian analysis is a theoretically grounded technique for combining information, learning from data and then forming predictions while representing uncertainties in these steps. Modern hierarchical Bayesian modeling has proved very effective in complicated settings. The strategy involves the specification of probability distributions, known as data models, for a variety of observational data sources. In addition, probability distributions, known as prior process models, are formulated for unknown processes interacting with each other on various space-time scales, as well as unknown parameters. The prior distributions for the processes are combined with the data models via Bayes' Theorem to produce a posterior probability distribution for the unknown quantities. This distribution serves as the basis for prediction. The reliance on probability means that uncertainty management is built-in to the approach. The key to our approach is to incorporate climate model output, at both global and regional scales, as if they are observations. These data are then combined with relatively simple, though physically based, prior models. The corresponding computations needed for prediction and uncertainty quantification are orders of magnitude easier than those associated with large-scale climate system models. Probability theory provides techniques for using the analyses to produce predictions and associated uncertainties of impacts. These results then serve as the basis for decision support. A critical challenge in current climate research is the prediction of regional climate behavior on decadal time scales. Reliable predictions of climate on decadal scales are the cornerstone of decision support for policy makers in their selection of adaptation strategies to address the impacts of climate change. Substantial information regarding climate is produced by climate system models. However, model results are subject to a variety of uncertainties. Further, these models are very demanding computationally and produce massive datasets, so there are severe limitations on our ability to quantify the uncertainties associated with model-based predictions. To respond to the challenge, we develop Bayesian techniques for predictive analyses that use climate model output in an efficient fashion and also deal with uncertainty. This combined use of large-scale climate system models, regional climate models, and simple climate models creates a synergistic environment in climate science. Next, because our strategy relies on comparatively simple calculations, we provide techniques for producing impact-specific and local or regional-specific predictive information in response to the needs of a variety of decision makers having a variety of responsibilities.