Data-driven and real-time prediction models for iterative and simulation-driven design processes

Sammanfattning: The development of more complex products has increased dependency on virtual/digital models and emphasized the role of simulations as a means of validation before production. This level of dependency on digital models and simulation togetherwith the customization level and continuous requirement change leads to a large number of iterations in each stage of the product development process. This research, studies such group of products that have multidisciplinary, highly iterative, and simulation-driven design processes. It is shown that these high-level technical products, which are commonly outsourced to suppliers, commonly suffer from a long development lead time. The literature points to several research tracks including design automation and data-driven design with possible support. After studying the advantages and disadvantages of each track, a data-driven approachis chosen and studied through two case studies leading to two supporting tools that are expected to improve the development lead time in associated design processes. Feature extraction in CAD as a way to facilitate metamodeling is proposed as the first solution. This support uses the concept of the medial axis to find highly correlated features that can be used in regression models. As for the second supporting tool, an automated CAD script is used to produce a library of images associated with design variants. Dynamic relaxation is used to label each variant with its finite element solution output. Finally, the library is used to train a convolutions neural network that maps screenshots of CAD as input to finite element field answers as output. Both supporting tools can be used to create real-time prediction models in the early conceptual phases of the product development process to explore design space faster and reduce lead time and cost.