Neural Operators for the Design of Acoustic Metamaterials

Neural operators have emerged as efficient, data-driven surrogates for modeling complex problems in physics and engineering. Unlike conventional neural networks, which approximate functions, neural operators learn entire operators, allowing them to predict solutions to parametrized partial differential equations without requiring retraining. This study explores the capability of neural operators to learn transmission loss curves of sonic crystal structures, using geometry parameters of a parametrized unit cell as inputs. Once trained, the neural operator facilitates fast predictions of transmission loss curves for corresponding metamaterial designs, eliminating the need for the computationally expensive meshing and solution procedures of conventional methods. Different neural operator implementations are evaluated against finite element reference solutions. The results demonstrate that all models achieve high accuracy and significantly outperform traditional methods in speed, advancing the possibilities of real-time metamaterial design.     «

Neural operators have emerged as efficient, data-driven surrogates for modeling complex problems in physics and engineering. Unlike conventional neural networks, which approximate functions, neural operators learn entire operators, allowing them to predict solutions to parametrized partial differential equations without requiring retraining. This study explores the capability of neural operators to learn transmission loss curves of sonic crystal structures, using geometry parameters of a paramet...     »