Physics-guided deep residual learning for sound absorption prediction of micro-perforated panels

Abstract To address the limited accuracy of analytical models, the high computational cost of finite element method (FEM) simulations, and the sample dependence of purely data-driven approaches for micro-perforated panel (MPP) sound absorption prediction, this study proposes a physics-guided deep residual network (PGDRN). A paired analytical-FEM dataset is constructed under identical structural parameters and frequency sampling points. Instead of directly fitting the complete FEM absorption curve, the FEM response is decomposed into an analytical baseline and an analytical-FEM residual correction. The network is trained to learn this residual discrepancy, while the analytical absorption coefficient and resonance mismatch are introduced as physics-guided features to enhance the representation of resonance-related frequency regions. Experiments on 1000 paired analytical-FEM samples show that PGDRN achieves an MAE of 0.0012, an RMSE of 0.0021, an R²of 0.9999, and a peak absorption coefficient error of 0.0022 on the test set. Compared with Mambular, the strongest data-driven baseline, PGDRN reduces MAE, RMSE, and peak error by 94.76%, 95.14%, and 95.53%, respectively. Ablation and limited-sample analyses confirm the contributions of residual learning and physics-guided features to prediction accuracy, peak-region stability, and sample efficiency. FEM sound-field visualization further supports the physical consistency of the predicted absorption behavior. The proposed method provides an accurate and interpretable surrogate model for rapid MPP sound absorption prediction under limited FEM samples.