Application of Machine Learning to Bridge Design

拟议研究计划的主要目标是开发一种基于机器学习的创新方法，使工程师能够有效地从先前建造的桥梁中提取知识，并无缝地应用这些知识，以支持增强的桥梁设计过程，该过程不仅适用于熟悉系统的设计，也适用于包含新元素的系统。机器学习的使用将为在设计中使用相关性创造一个更合理的基础，因此将为设计师提供一种方法来确定，在基础数据的数量和质量的作用下，相关性可以作为验证设计决策的基础的程度。该方法的第一步是应用可用的机器学习方法来支持复制设计过程，即由机器学习模型生成的设计反映用于训练模型的数据的特征。由于合适的数据集尚未存在，因此将创建一个数据集作为本研究计划工作的一部分。第二步是调整这些模型以支持创造性的设计过程，即由机器学习模型生成的设计包含原始数据中没有的特征。在此步骤中使用的策略将是使用足够数量的反映设计意图的附加条目来增加原始数据。通过这项工作要确认的假设是，相对少量的额外条目将足以重新训练模型，使它们能够支持具有这些新特征的桥梁设计。这一假设是基于桥梁结构系统之间的高度一致性，以及机器学习算法以极高的精度拟合高度复杂数据的能力。机器学习本质上是相关的，但桥梁设计的主要基础，需求和容量的检查，是基于因果关系，即物理原理的应用。该方法的第三步将是研究和描述基于相关性的机器学习生成的设计与基于因果关系建立的有效性之间的数学关系。提出的研究计划是机器学习在桥梁设计中的应用的第一个系统研究。它也是为数不多的研究之一，研究如何使机器学习（一种内在的复制技术）支持有意义的创造过程。因此，所提出的研究有可能为该领域的未来研究和发展设定方向，这将有可能在桥梁设计实践中带来重大转变。

The primary objective of the proposed research program is to develop an innovative methodology, based on machine learning, that will enable engineers to extract knowledge efficiently from previously constructed bridges and apply this knowledge seamlessly in support of an enhanced bridge design process that is adaptable not only to the design of familiar systems but also of systems that incorporate novel elements. The use of machine learning will create a more rational basis for the use of correlation in design and hence will provide a means for designers to determine, in function of the amount and quality of underlying data, the extent to which correlation can be relied on as a basis for validating design decisions. The first step in the methodology is to apply available methods of machine learning to support a replicative design process, i.e., one in which the designs generated by machine learning models reflect the features of the data used to train the models. Because a suitable dataset does not yet exist, one will be created as part of the work in this research program. The second step will be to adapt these models to support a creative design process, i.e., one in which the designs generated by machine learning models contain features that are not found in the original data. The strategy that will be employed in this step will be to augment the original data with a sufficient number of additional entries that reflect the design intent. The assumption to be confirmed through this work is that a relatively small number of additional entries will be sufficient to re-train the models to enable them to support the design of bridges with these new features. This assumption is based on the a high degree of consistency among the structural systems used for bridges, as well as the capacity for machine learning algorithms to fit highly complex data with great accuracy. Machine learning is intrinsically correlative yet the primary basis of bridge design, checks of demand and capacity, is based on causation, i.e., an application of physical principles. The third step in the methodology will be to investigate and characterize mathematically the relation between designs generated by machine learning on the basis of correlation and their validity as established on the basis of causation. The proposed research program is the first systematic study of the application of machine learning to bridge design. It is also one of the few studies of how to enable machine learning, an intrinsically replicative technique, to support a meaningful creative process. The proposed research thus has the potential to set the direction of future research and development in this field, which will has the potential to bring about a major transformation in the practice of bridge design.