CAREER: Cyberinfrastructure for Realizing Predictions with Uncertainty using Computational Modeling, Data, and Bayesian Inference

由于缺乏计算基础设施，无法将实验数据中的不确定信息整合到计算机模拟的复杂预测中，美国的科学和工程进步受到了限制。例如，化学模型的不确定性直接影响到美国能源和交通基础设施核心的高效燃烧系统的设计。将不确定性纳入计算模型的能力将直接促进对这种不确定性如何影响后续模型预测的理解，因此，我们基于这些复杂模型的预测做出可靠决策的能力。这项工作的目的是提供这样一个计算基础设施，可以很容易地利用部署在美国的大规模计算资源。这项工作中开发的框架可以通过提供确定大型不确定性来源的方法和工具来改进计算模型的开发。此外，该框架旨在通过促进识别最好地告知计算模型的实验场景，从而更好地设计物理实验。在这项工作中开发的网络基础设施还旨在为培训下一代科学家和工程师提供基础，使他们能够使用当代计算和数据支持的科学工具、理论和实践。因此，该项目与美国国家科学基金会促进科学进步和促进国家健康、繁荣和福利的使命是一致的。这项工作旨在开发的数字环境将支持基于偏微分方程（PDEs）的数学模型中未知参数的贝叶斯推断。具体来说，计算环境将促进不确定性计算预测的所有方面的创建、实验和检查：数学模型、偏微分方程的有限元公式、复杂系统的统计代理模型和解决统计逆问题的算法。在统计逆问题的解决和物理实验的设计方面，通过在复杂的模拟模型中容易地实现不确定性的存在和检查，将会取得重大进展。该平台将使用现代高性能计算算法，并可移植到极端规模的计算基础设施中，以解决这些模型。最后，所有元素都将部署在可移植、高效和易于使用的软件元素中，供整个社区使用。

The progress of science and engineering in the U.S. is now limited by the lack of computational infrastructure to incorporate uncertain information from experimental data into complex predictions made by computer simulation. For example, uncertainty in chemical models directly impacts the design of efficient combustion systems at the core of the U.S. energy and transportation infrastructure. The ability to incorporate uncertainty into computational models will directly advance understanding of how this uncertainty impacts subsequent model predictions and, therefore, our ability to make reliable decisions based on the predictions of these complex models. This work aims to provide such a computational infrastructure that can easily leverage the large scale computing resources deployed in the U.S. The framework developed in this work can enable improvements in the development of computational models by providing methodologies and tools for ascertaining the large sources of uncertainty. Additionally, the framework aims to enable better design of physical experiments by facilitating the identification of experimental scenarios that best inform the computational model. The cyberinfrastructure developed in this work is also aimed to provide the foundation for training the next generation of scientists and engineers to use contemporary computational and data-enabled science tools, theory, and practice. The project, thus is aligned to NSF's mission to promote the progress of science and to advance the national health, prosperity and welfare.The digital environment that this work aims to develop will support the Bayesian inference of unknown parameters in mathematical models based on partial differential equations (PDEs). Specifically, the computational environment will facilitate the creation, experimentation, and examination of all aspects of computational predictions with uncertainty: mathematical models, finite element formulations of PDEs, statistical surrogate models for complex systems, and algorithms for solving the statistical inverse problem. Significant progress in the state-of-the-art of both the solution of statistical inverse problems and the design of physical experiments will be made by easily enabling the presence and examination of uncertainty within complex simulation models. This platform will use modern high performance computing algorithms and be portable to extreme-scale computing infrastructure for the solution of those models. Finally, all elements will be deployed in portable, efficient, and easy-to-use software elements for use by the community at large.