Bayesian Methods For Climate Impact Uncertainty Quantification (Ref: 4386)

在我们的影响科学研究计划中，IBM Research正在探索如何通过量化极端气候事件的风险来使用人工智能来帮助建立对气候变化的适应能力。一个具体的重点领域是开发混合物理/数据驱动的气候影响模型，该模型既利用了传统的模拟，也利用了数量迅速增加的现场和遥感地球观测数据。到目前为止，研究中发现的一个关键差距是解决模型和数据在分辨率、领域、形式以及最重要的不确定性方面的异质性的适当方法。在构建和部署模型时，如果不能正确地组合和传播不确定性，可能会导致过度自信的预测，从而有可能被认为是不可靠和不准确的。博士项目将使用贝叶斯分层模型来解决这一差距，该模型将现实世界的气候(及其潜在影响)建模为一个潜在的、不可观察的过程，并将观测和模拟作为现实世界的不完美版本。这项研究的目的是使混合模型能够根据所有可用的信息来源，包括模拟模型和观测数据集，产生可靠和经过充分校准的高影响气候事件的概率。该方法将在IBM开发的实际应用中进行测试。主要工作内容包括：*开发和应用各种方法，以推断相关气候数据集和模型的错误结构和相互依赖关系(驻留在IBM PAINS数据库和相关的建模基础设施中)，使用参数统计推理，并在可能的情况下，使用专家知识；*开发和应用贝叶斯方法，以数学连贯的方式传播输入数据集和模型中的不确定性及其关系，以预测高影响的气候事件概率(如洪水、干旱、热浪)；*将上述开发的方法应用于缩小规模、区域化和将预测聚合到相关的空间和时间尺度；*扩展方法以评估复合气候风险，如热浪和洪水同时发生的可能性；*验证与IBM有关的现实世界商业应用中的决策，例如在供应链、金融服务、能源、公用事业和基础设施管理方面。

In our Impact Science research programme, IBM Research are exploring how AI can be used to help build resilience to climate change by quantifying the risk of extreme climate events. One specific area of focus is the development of hybrid physics/data-driven climate impact models, which leverage both traditional simulations and a rapidly-increasing volume of in-situ and remote-sensed earth observation data. A key gap identified in the research to date is an appropriate methodology to address heterogeneities across models and data in resolution, domain, form and, most significantly, uncertainty. Failure to correctly combine and propagate uncertainties when building and deploying models can result in over-confident predictions which are at risk of being considered unreliable and inaccurate. The PhD project will address this gap using Bayesian hierarchical modelling, which models the real world climate (and its potential impacts) as a latent, unobserved process, and observations and simulations as imperfect versions of this real world. The aim of the research is to enable hybrid models to produce reliable and well-calibrated probabilities of high-impact climate events, informed by all available sources of information, including simulation models and observational data sets. The methodology will be tested for real world applications in development at IBM.The main work components will consist of:* Development and application of methods to infer the error structures and interdependencies of relevant climate datasets and models (hosted in IBM PAIRS data repository and associated modelling infrastructure), using parametric statistical inference, and, where possible, expert knowledge;* Development and application of Bayesian methods to propagate uncertainties in input datasets and models, and their relationships, to predictions of high-impact climate event probabilities (e.g. flood, drought, heatwave), in a mathematically coherent way;* Application of the methods developed above to downscale, regionalise, and aggregate predictions to relevant spatial and temporal scales;* Extension of the methodology to assess compound climate risks such as the probability of co-occurrences of heat waves and floods; * Validation for decision-making in real world commercial applications relevant to IBM, for example, in supply chain, financial services, energy, utilities and infrastructure management.