Integrated numerical and machine learning framework for the vibro-acoustic analysis of car doors

Sound quality in modern vehicles is strongly influenced by the acoustic environment inside the cabin, where the car door plays a central role as both a vibrating structure and the mounting point for loudspeakers. This work introduces a novel integrated framework for car door vibro-acoustic design that combines numerical simulation, experimental validation, and machine learning. A simplified yet modular car door prototype is developed to reproduce the main structural and acoustic features of a real door while allowing flexible boundary conditions. A finite element model (FEM) of the prototype is validated through dedicated experimental campaigns, and is employed to generate a dataset of different structural configurations and their corresponding acoustic responses. This dataset is then used to train a feed-forward neural network aiming at predicting the sound pressure level with high accuracy at a negligible computational cost in comparison to FEM simulations. The trained model is finally used to perform statistical analysis of the effects of different boundary conditions in the radiated sound pressure. By identifying practical design guidelines for the optimal stiffness distribution of a general car door, this work demonstrates the value of integrating physics-based simulations with data-driven approaches. The results achieved can effectively support the design of complex vibro-acoustic systems, opening the way for more efficient optimization strategies in automotive engineering.