CDS&E: Efficient Uncertainty Analysis in Multi-physics Phase Field Models of Microstructure Evolution

非技术概述在材料-基因组倡议下，通过模拟具有许多可能的输入参数(成分、加工等)的材料，加速了材料设计。以便预测那些将产生所需性能的组合，例如，用于冷却应用的更高效的热电材料。在任何这样的设计周期中，输入中都存在不确定性，了解这些不确定性如何传播到输出特性中的不确定性是很重要的。对于复杂的计算，不确定性传播传统上是通过随机(“蒙特卡罗”)抽样来估计的，但当每次计算花费大量计算机时间时，这种抽样变得昂贵得令人望而却步。该项目的一个主要成果是开发并公开发布了有效的软件，以在计算材料科学计算中传播不确定性。虽然重点将放在材料微结构演变的模拟上，但拟议的框架将适用于与材料科学相关的广泛建模工具。还将开展活动以增加该项目的影响。这些活动包括建立一个丰富的二维和三维微结构资料库，该资料库可用作教授和开发微结构信息学新框架的模型系统；创建一个微结构动物园公民科学平台，以协助对该项目中创建的合成微结构进行标记和注释；以及为计算材料科学中的不确定性量化创建跨学科培训工具。该奖项还将支持两名研究生在材料科学和统计分析的界面上的培训。技术总结本研究的目标是为从模型输入到模型输出的不确定性传播创建顺序最优的抽样策略。这里考虑的问题是，当面对计算昂贵的模型，特别是热电和其他材料中微结构演化的相场模型时，如何进行准确而有效的不确定性传播。在本研究中，昂贵的模型是指在需要通过不确定性分析做出决定之前，资源将仅允许对其进行10阶模型评估的模型。这样的模型在许多领域都很流行，尽管这里的重点将放在计算材料科学的应用上。具体地说，在这个项目中，研究小组将展示不确定性在CALPHAD相场模型链中的有效传播，该模型链试图描述材料在化学和弹性驱动力下的微结构演变。这些模型往往是高度非规则的，因为模型输出可以根据被采样的输入/参数空间中的区域而定性地不同。此外，它们的计算成本很高，全三维模拟需要超过10,000个CPU小时。此外，输入/参数空间往往是高维的，具有20多个随机输入条件和模型参数。这种不确定性传播框架的主要特征之一是它是非侵入性的，因为它在模型的输入空间上工作。因此，所需要的唯一信息是输入参数的联合概率分布，这使得该框架具有广泛的适用性，超出了开发该框架所使用的特定测试问题(S)。该框架能够对先前执行的模型评估进行重新加权，以最大限度地提高不确定性传播的效率，而额外的计算开销可以忽略不计。PI将在开放许可下通过Github存储库发布不确定性传播和相域代码。他们将建立一个公共材料模型和数据管理系统(MMDMS)，使他们在研究过程中产生的相场和DFT数据广泛可用，并与其他数据存储库协调。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。

Nontechnical SummaryMaterials design has been accelerated under the Materials-Genome Initiative by simulating materials with many possible input parameters (composition, processing, etc.) in order to predict those combinations that will yield desired properties, for example, more efficient thermoelectrics for cooling applications. In any such design cycle, there are uncertainties in the inputs, and it is important to understand how these uncertainties propagate to uncertainties in the output properties. For complex computations, uncertainty propagation has traditionally been estimated with random ("Monte-Carlo") sampling, but when each calculation takes substantial computer time, such sampling becomes prohibitively expensive. A major outcome of this project is the development and public release of efficient software to propagate uncertainty across computational materials-science calculations. While the focus will be on simulations of the evolution of materials microstructures, the proposed framework will be applicable to a wide range of materials-science-relevant modeling tools. Activities will also be conducted to increase the impact of the project. These include the generation of an extensive library of 2D and 3D microstructures that can be used as model systems to teach and develop new frameworks for microstructure informatics, the creation of a microstructure-zoo citizen-science platform to assist in labeling and annotating synthetic microstructures created in this project, and the creation of interdisciplinary training tools for uncertainty quantification in computational materials science. This award will also support the training of two graduate students at the interface of materials science and statistical analysis.Technical SummaryThe objective of this research is to create sequentially optimal sampling policies for the propagation of uncertainty from model inputs to model outputs. The problem considered here is that of conducting accurate and efficient uncertainty propagation when faced with computationally expensive models, specifically phase-field models of microstructure evolution in thermoelectrics and other materials. Expensive in this research refers to models for which resources will only permit order 10 model evaluations prior to the need to make a decision informed by the uncertainty analysis. Such models are prevalent in many fields, though the focus here will be on computational materials-science applications. Specifically, in this project, the research team will demonstrate the efficient propagation of uncertainty across CALPHAD-Phase Field Model chains that attempt to describe the microstructure evolution of materials under chemical and elastic driving forces. These models tend to be highly non-regular in that the model output can differ qualitatively depending on the region in the input/parameter space being sampled. Moreover, they are computationally expensive, with full three-dimensional realizations of the simulations requiring upwards of 10,000 CPU-hours. In addition, the input/parameter space is often high dimensional, with more than 20 stochastic input conditions and model parameters. One of the main features of this uncertainty propagation framework is that it is non-intrusive, as it works on the input space of the models. Thus, the only information needed is the joint probability distribution of the input parameters, which makes the framework widely applicable, beyond the specific test problem(s) used to develop it. The framework is capable of reweighting previously executed model evaluations in order to maximize efficiency for uncertainty propagation with negligible additional computational expense. The PIs will release uncertainty-propagation and phase-field code through a Github repository under open licenses. They will set up a public Materials Models and Data Management System (MMDMS) to make the phase-field and DFT data they generate in the course of this research widely available and coordinate with other data repositories.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.