Enhancing Identifiability of Computer Simulation Models via Design for Calibration

该奖项的研究目标是开发一种校准设计（DfC）方法，以提高仿真模型的可识别性。可识别性是指在预测建模中分离两种一般不确定性来源-参数不确定性和模型不确定性的难度。参数不确定性是由于对基本物理参数的不完全了解造成的，模型不确定性是由于模拟模型中的近似值和其他不准确性造成的。为了确保适当的识别这些不确定性时，结合丰富的模拟数据与有限的物理实验数据，研究利用两个关键的有利因素，第一个是固有的多响应性质的计算机模拟，第二个是广泛的模拟结果的可用性之前，设计的物理实验。在贝叶斯框架内使用空间随机场建模，该方法将确定实验测量的响应变量的最佳子集和在实验中使用的输入设置的最有效组合，其目标是最佳地增强不确定性的可识别性。如果成功的话，这项研究的结果将有助于确保预测不确定性的可识别性，从而使有限的实验资源得到最有效的利用。这在所有工程学科的基于仿真的工程和科学中至关重要，原因有很多，不仅仅是实现良好的近视预测。学习和区分真实的物理参数和仿真模型的不准确性对i）比实验测试台复杂得多的新产品/过程设计，ii）改进未来几代仿真代码，以及iii）在更广泛的输入区域集合上提供更准确的预测具有广泛的影响。由于该方法不依赖于特定类型的仿真代码或应用领域，因此预计将具有广泛的适用性。这项工作将利用西北大学跨学科博士预测科学与工程设计集群的广泛支持，通过该集群，将绘制多学科试验台应用程序，并将结果传播到不同的工程领域。

The research objective of this award is to develop a Design for Calibration (DfC) methodology for enhancing identifiability in simulation models. Identifiability refers to the difficulty in separating two general sources of uncertainty - parameter uncertainty and model uncertainty - in predictive modeling. Parameter uncertainty results from imperfect knowledge of the underlying physical parameters, and model uncertainty results from approximations and other inaccuracies in a simulation model. To ensure proper identifiability of these uncertainties when combining abundant simulation data with limited physical experimental data, the research takes advantage of two key enabling factors, the first being the inherently multi-response nature of computer simulations, and the second being the availability of extensive simulation results prior to designing the physical experiment. Using spatial random field modeling within a Bayesian framework, the methodology will determine the best subset of response variables to measure experimentally and the most efficient combination of input settings to use over the experiments, with the objective of optimally enhancing identifiability of the uncertainties. If successful, the results of this research will help ensure identifiability of predictive uncertainties in a manner that allows limited experimental resources to be used most efficiently. This is critically important in simulation-based engineering and science across all engineering disciplines for many reasons that extend beyond achieving good myopic prediction. Learning and distinguishing the true physical parameters and simulation model inaccuracies has broad-reaching implications for i) new product/process designs that are much more complex than the experimental testbed, ii) improving future generations of simulation code, and iii) providing more accurate prediction over a wider set of input regions. Because the methodology is not tied to a particular type of simulation code or application domain, it is expected to be widely applicable. This work will leverage the broad-based constituency of the interdisciplinary doctoral Predictive Science and Engineering Design cluster at Northwestern, through which multidisciplinary testbed applications will be drawn and the results disseminated throughout different engineering domains.