Project Details
Projekt
AI-based optimization of vehicle crashworthiness design: Taking randomness out of design optimization
Applicant
Professor Dr.-Ing. Marcus Stoffel
Subject Area
Engineering Design, Machine Elements, Product Development
Mechanics
Mechanics
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 501877598
Design optimization is a core aspect of component design. The quest for optimized design is of great importance now more than ever, due to depleting natural resources and environmental concerns. Though, there exist a catalogue of classical optimization methods, they are computationally expensive, are less scalable, and the most important being that the results obtained are sub-optimal and have lower sample efficiency. Therefore, in this study, we propose machine learning-based (ML) methods to address these issues with design optimization of crashworthiness as the core optimization problem. Vehicle crash simulations involve highly nonlinear deformation patterns and are computationally intensive and therefore a suitable example to demonstrate the effectiveness and the efficiency of the proposed framework. The proposed methods are aimed to address the randomness involved in the methods available in the literature and gear towards a robust, more generalizable and sustainable solution. The framework comprises multiple design assistants, each of which can operate individually for different optimization tasks, working in tandem to perform the design optimization. The design assistants include a novel Graph neural network (GNN) based FE-surrogate, a novel Graph-based intelligent finite element solver, an evolutionary reinforcement learning (ERL) model, and a hardware inference module together to investigate multidimensional optimization of vehicle crashworthiness design. The AI inference module enables a seamless plug and playable design optimizer which can be used on any workstation without the need administrator level access or new program packages. The framework is then extended to a wide range of optimization problems in the literature to demonstrate the robustness of the proposed design assistants and their generalizability. The study aims to modernize, optimize and approach the optimization problem in a sustainable way, reducing the overall computational effort, improving accuracy compared to existing methods and save energy through custom inference chips.
DFG Programme
Priority Programmes
