Convolutional neural networks to predict dispersion surfaces-based properties of acoustic metamaterials with arbitrary-shaped unit cells

• A deep learning framework for the analysis and design of acoustic asymmetric two-dimensional metamaterials. • High predictive accuracy ( R 2 > 0.99) through optimized hyperparameters and model design. • A GUI implementation available on GitHub ensuring reproducibility and supporting future development. • Bandgap properties with high accuracy, validated through inverse design. • New insights into how geometric features influence bandgap widths to improve materials design strategies. This study presents a deep-learning framework for predicting and optimizing the bandgap properties of 2D acoustic metamaterials with arbitrary-shaped unit cells. The convolutional neural network (CNN) presented in this work achieves exceptional accuracy, with coefficients of determination ( R 2 ) exceeding 0.99, in predicting both bandgap widths and positions. We improve model interpretability using SHapley Additive exPlanations (SHAP), revealing how geometric features, including concave and convex curvatures, impact bandgap properties. Additionally, an inverse design process utilizing particle swarm optimization technique identifies metamaterial geometries that meet specific bandgap ratio targets, with prediction errors consistently below 0.25 %. The implementation code is available on GitHub, ensuring reproducibility and facilitating future research. This framework offers a powerful tool for designing and optimizing next-generation acoustic metamaterials.