Artificial intelligence-driven surrogate modelling and optimisation for hot-stamped safety-critical automotive components
File(s)
Author(s)
Zhou, Haosu
Type
Thesis
Abstract
The aim of this study is to establish a comprehensive framework for developing and evaluating advanced Artificial intelligence-driven surrogate models (AISM) with structured data representations, tailored for industrial applications in metal stamping, particularly hot stamping. This research addresses critical challenges in achieving accurate and efficient simulations for complex geometries by exploring the capabilities of image-based and graph-based AISMs. Case studies and evaluation criteria were defined for performance assessment in terms of accuracy, scalability, and computational efficiency.
The research focuses on: (1) validating image-based AISM for cold stamping simulations with simplified geometries, comparing its performance against traditional scalar-based models; (2) evaluating the potential of image-based AISM for hot stamping with limited data to determine whether large datasets are essential for effective training; (3) extending image-based AISM to real-world geometries and integrating it with a differentiable shape generator for real-time, non-parametric shape optimisation; and (4) proposing an efficient, scalable graph-based AISM architecture based on graph convolutional operations but also integrates the advantages of images, aiming to address challenges faced by existing graph-based surrogate models.
Overall, the findings underscored the advantages of structured data representations, because both image-based and graph-based AISMs have demonstrated superior accuracy and scalability compared to scalar-based approaches. The findings related to aspect (2) revealed that even with small data, AISM with structured data representations can achieve high accuracy, challenging the conventional need for extensive datasets in industrial applications. The findings related to aspect (3) demonstrated the effectiveness of the proposed non-parametric shape optimisation framework, positioning it as a viable tool for improving design efficiency and precision in metal stamping applications. The findings related to aspect (4) provided a methodology of surrogate modelling with integrated advantages of images and graphs. These research findings provided insights for future development of surrogate modelling and optimisation, aiming for better efficiency, accuracy and generalisability.
The research focuses on: (1) validating image-based AISM for cold stamping simulations with simplified geometries, comparing its performance against traditional scalar-based models; (2) evaluating the potential of image-based AISM for hot stamping with limited data to determine whether large datasets are essential for effective training; (3) extending image-based AISM to real-world geometries and integrating it with a differentiable shape generator for real-time, non-parametric shape optimisation; and (4) proposing an efficient, scalable graph-based AISM architecture based on graph convolutional operations but also integrates the advantages of images, aiming to address challenges faced by existing graph-based surrogate models.
Overall, the findings underscored the advantages of structured data representations, because both image-based and graph-based AISMs have demonstrated superior accuracy and scalability compared to scalar-based approaches. The findings related to aspect (2) revealed that even with small data, AISM with structured data representations can achieve high accuracy, challenging the conventional need for extensive datasets in industrial applications. The findings related to aspect (3) demonstrated the effectiveness of the proposed non-parametric shape optimisation framework, positioning it as a viable tool for improving design efficiency and precision in metal stamping applications. The findings related to aspect (4) provided a methodology of surrogate modelling with integrated advantages of images and graphs. These research findings provided insights for future development of surrogate modelling and optimisation, aiming for better efficiency, accuracy and generalisability.
Version
Open Access
Date Issued
2024-12-31
Date Awarded
2025-06-01
Copyright Statement
Attribution-Non Commercial-No Derivatives 4.0 International Licence (CC BY-NC-ND)
Advisor
Li, Nan
Childs, Peter
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)