A neural network–driven response surface surrogate with virtual DOE for interpretable vehicle interior sound quality modeling

This study presents a hybrid framework integrating a GA-BP neural network with a second-order Response Surface Methodology (RSM) surrogate for vehicle interior sound quality modeling. Five acoustic indicators, including A-weighted SPL, loudness, sharpness, roughness, and articulation index (AI index), are mapped to subjective sound quality scores. The GA-BP model was evaluated using repeated nested cross-validation, while a neural network-driven virtual Design of Experiments (DOE) was used to generate structured query points for explicit RSM construction. Additional vehicle measurements at 55 and 65 km/h showed close agreement between model predictions and subjective ratings, providing consistency checks within the investigated operating range. Response surface analysis revealed SPL-related changes in local sensitivity and conditional interactions among acoustic parameters. The proposed framework combines nonlinear prediction with explicit surrogate interpretation, providing an efficient approach for vehicle sound quality analysis under limited experimental data.