Deep learning-based prediction and interpretability of physical phenomena for metaporous materials

Abstract

In this paper, we propose a fast, accurate, and interpretable deep learning (DL)-based method for predicting and interpreting the sound absorption characteristics of metaporous materials. A novel deep convolutional neural network (CNN) model that hybridizes the physical knowledge of metaporous materials, i.e., resonances, is presented to accurately predict the sound-absorbing performances of metaporous materials based solely on their geometric information. Our proposed DL model inspired by locally and globally resonant mechanisms is trained to effectively capture the individual geometric information of embedded split-ring resonators and reflect the global structural interaction between the resonators and their surrounding environments, such as a porous material and a hard-backing layer. Both quantitative and qualitative results demonstrate that the proposed method outperforms the comparative methods, achieving far more accurate predictions of sound absorption coefficients with an average frequency-wise absolute difference of 0.009 and R2 score of 0.98. Besides, the average computation time per single case of the proposed method is observed to be 708 times faster than that of the existing method. Further, we examine the possibility that our physics-inspired proposed model can derive physical relevance appearing in metaporous materials, showing that physically aligned interpretability can be virtually obtained by the proposed approach in the form of visual activation maps. Our study delivers the potential to address various acoustic metamaterials challenges that require real-time, accurate, and interpretable analyses.