Application of a Machine Learning-Based Topology Optimization Method on an Automotive B-Pillar for Crashworthy Lightweight Design

Abstract

This work investigates the application of voidML to a reinforcement patch of a B-pillar which is subject to a side impact load case. voidML is a machine learning-based topology optimization method that iteratively inserts voids into crash-loaded thin-walled structures while preserving the deformation behavior of a reference design. Void placement is guided by a Gradient Boosted Decision Trees (GBDT) model trained on local state features extracted from crash simulations, which predicts the Displacement Field Deviation (DFD) for each candidate void position. Multiple competing designs are pursued in parallel through a design tree for better design space exploration. The mass of the reinforcement patch is reduced from 2.413 kg to 2.141 kg, corresponding to a mass reduction of 11.3 %, within 48 iterations and a total of 191 crash analyses. The force-intrusion curves of the reference design and the final design remain close to each other, confirming that the DFD constraint effectively controls the deviation from the reference design. The results indicate the potential of voidML for lightweight design of crash-loaded thin-walled structures in industrial applications.