Beschreibung:
When tailoring porous absorbers in acoustic applications, an appropriate acoustic material model, as well as the relationship between the material model parameters and the microscale geometry of the material, is indispensable. This relationship can be evaluated analytically only for few simple material geometries. Machine-learning models can close this gap for complex materials, but due to their black-box nature, the interpretability of obtained inferences is rather low. Therefore, an existing neural network model that predicts the acoustic properties of a porous material based on the microscale geometry is subject to statistics-based sensitivity analysis. This is conducted to gain insights into the relationship between the microscale geometry and the acoustic material parameters of a generic bar-lattice design porous material. Although it is a common approach in the field of explainable artificial intelligence research, this has not been widely investigated for porous materials yet. By deriving statistics-based sensitivity measures from the neural network model, the explainability and interpretability is increased and insights into the relationship of the acoustic properties and their microscale geometry of the porous specimen can be obtained. The results appear plausible and comparable to existing studies available in the literature, showing if and how the bar-lattice geometry influences the acoustic material parameters. Moreover, it could be shown that the applied global sensitivity analysis method allows us to not only derive a one-to-one parameter impact relation, but also reveals interdependencies that are important to address during a material tailoring process.