Data Engineering for Data-Driven Design

Doctoral thesis, 2025

Today, the design process of products heavily uses modeling and simulation to assess how well they fulfill the requirements. Iterative and simulation-driven design processes have reduced testing costs but not the lead time associated with them. One example is the airbag design process, which involves hundreds of manual and digital prototyping loops before physical tests are conducted. Addressing these issues requires more efficient design evaluation. Data-driven design has shown significant potential in accelerating design cycles; however, it is also hindered by its reliance on the availability of data. The product design process establishes parameterization conventions that must be followed when analyzing. Running large-scale simulations to generate labels in design datasets can be costly, often leading to underexplored design spaces. The datasets generated are created individually without a system perspective, which often restricts future design changes. This thesis investigates data engineering as a critical driver of data-driven design to support design evaluation. The goal is to identify and mitigate the main problems associated with data generation by exploring new methods for constructing, extracting, and organizing features and labels. The thesis aims to streamline digital verification cycles and reduce engineering design lead time. Extracting features from alternative geometric representations, such as the medial axis, is proposed to reduce the reliance of data-driven evaluation methods on model parameterization. Using these parameters is shown to be superior to CAD parameterization for predictive tasks. Therefore, the concept of sleeping parameters is suggested as a potentially impactful feature in the dataset, which enhances knowledge encapsulation and transfer within the product structure. Image regression is proposed using design screenshots as an alternative method for building prediction models in design evaluation. The necessary large-scale dataset for this task is created through dynamic relaxation. To address design change analysis in design evaluation, the sleeping parameters concept is used in a framework called Product Dataset Platform, where component-level data is leveraged for system-level evaluations. It was shown that these solutions, individually, enable the rapid exploration of novel design configurations and accelerate the validation process. The results contribute to the foundation for CAD-CAE integration by providing a design evaluation framework. The suggested supports enable predictive capacity that can help map function to form and allow more efficient evaluation after the design change. The findings show a way to a data-driven design evaluation method that accelerates design iterations and reduces engineering design lead time.