<p>This paper introduces voidML, a novel machine learning-based method for topology optimization of crashworthiness structures that addresses the application of inserting voids (cutouts) into thin-walled structures. The goal is to insert voids into the structure while preserving the deformation behavior of the reference design, which is assumed to already satisfy the applicable crash functional requirements. The method is therefore ideally suited for optimizations in late phases of the vehicle development process. The structural impact of void insertions is predicted using a Gradient Boosted Decision Trees (GBDT) model trained on local state features extracted from crash simulations. The GBDT estimates the Displacement Field Deviation (DFD), a scalar metric quantifying changes in structural behavior, enabling efficient void placement without initial sampling during topology optimization. Predictions can be made on unseen crash load cases and components, as the GBDT aims to learn general patterns using local states and is never exposed to the entire component. During optimization, competing designs are pursued. This reduces the probability of getting trapped in local minima. voidML is applied to two crash load cases, reducing the mass iteratively by inserting voids while maintaining the DFD below a prescribed constraint. Several areas of potential future work are identified that could lead to further improvement.</p>