Abstract
The challenge for large-scale implementation of Multidisciplinary Design Optimization (MDO) algorithms in solving real-world automotive structural design problems is the time associated in running huge Finite Element Models that involve solving highly non-linear phenomenon (such as crash) involving large number of design variables and multiple performance constraints pertaining to various domains (NVH, Durability and Crash). With the availability of powerful Computer Aided Engineering (CAE) tools, it has become possible to generate the data of performance responses for a huge number of design points considering all the design variables that exist in a real-world design problem. Traditional Response Surface Method based approaches fail to handle such huge data and therefore, it is important to explore the Machine Learning based approaches that can easily handle huge and complex real-world data. The present work focusses on developing a methodology for lightweighting of an automotive structure using machine learning based, MDO. In this study, the objective is to minimize the mass of the front end structural components, of a passenger car that are likely to effect the vehicle crashworthiness and NVH performances. Size optimization is performed by considering the gages of various components as the design variables and front-end intrusions during an IIHS offset impact test simulation and the modal frequency of a critical structural member as the constraint variables. Two different methods, namely, neural network based and the random forest based machine learning algorithms were used to develop predictive models for use in design optimization. It is shown that the data mining based optimization methods are substantially more efficient and applicable in solving the real-world vehicle design problems that contain many design variables. A considerable weight saving of nearly 25 % with respect to baseline design is obtained in the present study.