This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Machine learning (ML) is emerging as a key tool for predicting crash dynamics in near real-time, driving significant advancements in automotive safety engineering. In this domain, high-fidelity synthetic data, such as finite element (FE) crash simulation data, plays a pivotal role. Real-world crash data often presents challenges such as limited availability, noise, and incomplete information, making it difficult to train robust ML models. By contrast, simulation data provides detailed insights into the underlying physical phenomena, offering a controlled environment to generate diverse datasets. This makes simulation data an indispensable resource for the development of predictive models that aim to improve vehicle safety and occupant protection. Traditional ML approaches in this field often rely on scenario-level input parameters, such as impact velocity, and collision angles, to predict outcomes like intrusion or injury levels. While these approaches are effective to a degree, they frequently fall short of leveraging the rich, granular information embedded within simulation data. This limitation can result in suboptimal predictive accuracy, particularly when dealing with complex crash dynamics involving multiple interacting factors. A domain knowledge-guided methodology is introduced to address this limitation, segmenting the problem domain into smaller, homogeneous sub-domains based on specific physical phenomena. By reducing the required ML model complexity within each sub-domain, tailored ML models are developed to optimize predictions for specific crash dynamics, enhancing both data efficiency and prediction model accuracy. This approach leverages hierarchical model trees informed by domain expertise, ensuring that each sub-domain's unique traits are effectively captured without relying on a singular, overly complex model. Embeding domain knowledge through segmentation not only allows for the customization of sub-model architecutres but also facilitates adaptive data generation. The proposed methodology improves predictive performance and reduces training sample requirements while preserving fidelity. Comparative evaluations demonstrate superior accuracy and robustness in capturing nuanced relationships within crash simulation data, positioning this approach as a significant step forward in the application of ML to safety-critical domains.