Deep learning-based approach for linking microstructural and macroscopic acoustic properties of sound-absorbing polyurethane foam

Sound absorption in porous materials is fundamentally governed by their microstructural morphology yet establishing a quantitative and design-oriented relationship between microstructure and acoustic behavior remains challenging. To address this challenge, a deep learning-based acoustic modeling framework is proposed for analyzing the microstructure of flexible polyurethane (PU) foam and predicting its acoustic performance. A microscopic analysis model is developed to semantically segment SEM images using a U-Net model and quantitatively extract the distribution parameters of microstructural properties, including cell size, pore size, pore shape factor, and strut thickness. An artificial neural network model is developed to model the relationship between these microstructural parameters and acoustic performance measured using an impedance tube, based on 210 flexible PU foam samples including thermally aged and non-aged materials from multiple manufacturers. Finally, the proposed approach is validated through comprehensive performance evaluation, comparison with experiments and alternative methods, and analysis of microstructural parameter contributions. The validated framework establishes a quantitative link between microscale morphology and acoustic performance, providing data-driven acoustic insights and practical guidance for acoustic material design, including feature selection, optimization of acoustic performance, and fabrication of sound-absorbing materials for applications such as automotive and construction.