Foundation Models Orchestrate AI for Pedestrian Protection Design

AI-driven engineering workflows face unique challenges in crash safety design, particularly due to the highly nonlinear contact dynamics, material properties, and discrete state transitions involved in crash events. These complexities make it difficult for data-driven surrogate models to accurately capture the physics. Researchers have introduced what they believe to be the first foundation model-orchestrated workflow specifically for crash safety design, enabling surrogate-assisted exploration for pedestrian protection. This innovation dramatically cuts evaluation time from hours per CAE simulation to mere seconds. The workflow integrates four key components. First, a surrogate model is trained on CAE crash simulations to predict pedestrian leg injury metrics based on design parameters, achieving high accuracy and providing reliable prediction intervals. Second, a multiobjective evolutionary search algorithm (NSGA-II) is employed to discover diverse sets of feasible design parameters that meet user-specified constraints. Third, a morphing-based geometry generator translates these parameters into topology-preserving 3D shapes. Finally, a natural-language interface, powered by a Large Language Model (LLM), orchestrates the entire workflow, while a vision-language model assists in the semantic comparison of the generated designs. In a practical case study involving an automotive front-bumper design, this workflow successfully generated 35 distinct safety-compliant alternatives from a single exploration. This process, which would typically take weeks using conventional CAE iteration, was completed in a fraction of the time. These results highlight the potential of foundation models to act as crucial integration layers between machine learning surrogates and physics-based simulations, thereby extending AI capabilities into safety-critical engineering domains.