The protection of occupants and other road users is an important aspect in modern vehicle development, requiring various crash tests across global markets. However, physical crash tests are inherently variable, leading to variations in test outcomes. To address this, virtual simulations are increasingly used to assess the robustness of occupant safety systems. Modern strategies leverage machine learning models for fast evaluation, enabling the optimization and in-depth analysis of occupant protection under diverse crash scenarios. This thesis investigates how geometry information of crash simulation models can be effectively integrated into supervised machine learning models to predict outcomes in crash and occupant safety scenarios. Specifically, the initial geometric data is used to predict the maximum chest deflection caused by the seatbelt on an Anthropometric Test Device (ATD) during crash simulations. To achieve this, two datasets are evaluated, where variations in the initial seating position of the ATD and seatbelt placement on the ATD resulted in different kinematics and loadings during the simulations. State-of-the-art deep learning models from the 3D data processing domain are employed, utilizing various representation methods, including point clouds, voxels, and graphs. The findings demonstrate that data preprocessing is crucial for effectively integrating 3D data into deep learning models. It is found that a global normalization approach — combining all point clouds into a single large point cloud before applying unit sphere normalization — better preserves inter-sample distances, leading to increased regression performance. Additional techniques, such as downsampling to 1024 points using farthest point sampling, and data augmentation through random scaling, also contribute to significant performance improvements. Overall, the deep learning models successfully capture meaningful geometric features, often outperforming baseline models that rely on domain-derived geometric features. Additionally, the exploration of baseline models’ explainability using Shapley values indicate that geometric features hold greater importance than system features in determining regression accuracy, further emphasizing the central theme of this thesis. Furthermore, an analysis of the saliency maps from the deep learning models reveals that these models make predictions based on only a few key regions of the input geometries, similar to the process of extracting geometric features. This thesis serves as a proof of concept, demonstrating the value of integrating 3D data directly into machine learning models in the context of vehicle safety, rather than relying solely on extracted geometric features. The findings suggest that to achieve optimal predictive performance, enriching 3D data with extracted geometric features is the most effective approach. The concepts and insights presented in this thesis can also be applied to other use cases. Future research could leverage these methods to predict additional dummy loadings in various vehicle safety scenarios.