Data-driven optimization for enhanced computational engineering design

数据驱动的科学一直在为从生物学到金融学的几个不同学科的进步做出贡献。在工程学方面，这重新引起了人们对机器学习和人工智能的兴趣，特别是在网络物理系统的背景下，这些系统的开发旨在支持“工业4.0”、智能交通系统或“智能”医疗系统，仅举几个例子。共同的主题是以数据的形式获取和分析信息，以更好地理解、建模和预测这些协作的网络-物理系统的行为，从而改进它们的设计。这种“数字孪生”模式旨在利用实时数据不断更新物理实体的计算模型，以(重新)优化它们的设计、运行、维护、维修或更换等，这些丰富的上下文数据的可用性可以通过优化手段显著提高计算工程设计过程。与此同时，出现了几个阻碍传统优化方法应用于这一新范式的挑战，例如：为了有效，数字孪生兄弟很可能是高维的；此外，它们通常是黑箱：梯度信息通常要么不可用，要么几乎不可能可靠地逼近。数据集可以非常大和/或稀疏，并且可以包括不连续和/或离群值。而且，它们是连续追加的。在设计变量和参数所跨越的输入空间的不同区域，特别是在频繁的数据更新的情况下，模型的预测能力所需的充分性可能会有很大差异。必须捕获和协调连接的系统之间的交互，以确保协作网络无缝集成和可互操作。考虑到网络物理系统之间既存在物理链接又存在计算链接，这一点尤其具有挑战性。拟议的研究计划的目标是通过在协作网络物理系统的背景下开发一个数据驱动的计算工程设计优化框架来应对上述挑战。为了实现这一目标，我们将开发和集成一个数据驱动的环境，用于自适应的、基于充分性的多模型管理和验证，以及用于分布式系统的严格的免导数优化算法和协调技术。拟议的研究计划将培养5名博士生、3名硕士和5名本科生，为下一代工程和生产系统做好准备。

Data-driven science has been contributing to the advancement of several diverse disciplines ranging from biology to finance. In engineering, this has generated renewed interest in machine learning and artificial intelligence, especially in the context of cyber-physical systems that are developed to enable “industry 4.0,” intelligent transportation systems, or “smart” healthcare systems, just to name a few examples. The common theme is the acquisition and analysis of information in the form of data to better understand, model and predict the behaviour of these collaborative cyber-physical systems so that their design can be improved. This “digital twin” paradigm aims at continuously updating the computational models of their physical counterparts using real-time data to (re-)optimize their design, operation, maintenance, repair or replacement, etc.The availability of such rich contextual data can enhance the computational engineering design process significantly by means of optimization. At the same time, several challenges arise that hinder traditional optimization methods from being applied to this new paradigm, e.g.: To be effective, digital twins are likely to high dimensionality; moreover, they are typically blackboxes: gradient information is typically either not available or almost impossible to approximate reliably. Data sets can be quite large and/or sparse, and can include discontinuities and/or outliers. Moreover, they are appended continuously. The required adequacy of the predictive capability of the models can vary significantly in different areas of the input space spanned by design variables and parameters, especially in light of frequent data updates. The interactions among the connected systems have to be captured and coordinated to ensure that the collaborative network is seamlessly integrated and interoperable. This is especially challenging considering that there exist both physical and computational links among the cyber-physical systems.The objective of the proposed research program is to address the above challenges by developing a framework for data-driven computational engineering design optimization in the context of collaborative cyber-physical systems. To accomplish that, we will develop and integrate a data-driven environment for adaptive, adequacy-based multi-model management and validation with rigorous derivative-free optimization algorithms and coordination techniques for distributed systems. The proposed research program will train 5 PhD students, 3 Masters students, and 5 undergraduate students, preparing them for the next generation of engineering and production systems.