Abstract
Beams are used in the doors of cars to provide adequate strength and stiffness against side impact. The door beam needs to satisfy the FMVSS214s regulatory requirements and meet other associated performance constraints such as not to undergo buckling at the beam ends. Conventional surrogate models fail to give accurate predictions of such qualitative performance constraints. The aim of this study is to explore capability of various machine learning models to predict qualitative and classificatory responses such as end-buckling in door beams and then integrate this prediction capability to come up with an optimization framework for such door beams. Various sections of the beam are identified as gauge and shape variables and parametrized appropriately. The Strength Two Orthogonal Array (STOA) technique is used to generate the DOE. LS-DYNA® simulations are carried out to evaluate performances for DOE points. Further, one portion of the data is also used as a training set for different Machine Learning models and the other portion is used for verification. Different machine learning methods are explored for their efficacy in predicting beam end-buckling, which is a typical qualitative and classificatory response. Random Forest based ensemble classifier is found to outperform others in prediction accuracy. This model is tuned further to enhance accuracy of prediction and subsequently integrated with Particle Swarm Optimization (PSO) technique to come up with the framework for door beam optimization. This tool is shown to result in substantial decrease of the door beam mass while maintaining or improving all the performance constraints (qualitative and quantitative)