4 June 2024 in Building and Environment
21 August 2025 in Automation in Construction
6 July 2026 in SSRN Electronic Journal
TETomás Méndez Echenagucia
Ventilated acoustic barriers can reduce noise while maintaining airflow, but their design is highly sensitive to geometry. This study presents a surrogate modeling framework for predicting the frequency-dependent transmission loss of horn-like helical ventilated acoustic barriers with Fano-like resonance behavior. A simulation dataset was generated using COMSOL over a Sobol-sampled design space. Instead of using only design-specific variables, the models were trained with physics-informed features that combine local spatial descriptors with global propagation quantities, including direct and indirect path lengths and areas. Three surrogate models were evaluated: CatBoost, a one-dimensional convolutional neural network, and a multilayer perceptron. The models were tested on training, test, and extrapolation sets to assess both interpolation and generalization. All models achieved strong test performance, with overall 𝑅2 values of 0.921, 0.939, and 0.927 for CatBoost, CNN, and MLP, respectively. Extrapolation performance decreased for all models, indicating the difficulty of predicting designs beyond the boundary of the sampled design space. Feature-importance and design-space analyses showed that global propagation descriptors, especially path-length and area-based features, strongly influenced model predictions. The proposed framework enables efficient acoustic response prediction and provides a basis for future optimization of ventilated acoustic metamaterials