Revealing Structure-Property Linkages using Explainable AI

The development of a new vehicle concept at the DLR (German Aerospace Centre) involves the use of high-performance sandwich composites that combine low weight with high energy dissipation. Among these materials, sandwich composites with a polyurethane foam core and a metallic face material show outstanding performance. However, the design of these composites requires knowledge of the mechanical properties of the constituent materials. For solid metallic materials the mechanical quantities are generaly known. But for composite and open-cell porous foams, the properties are unknown as they depend on the composition and structural characteristics. Experimental determination requires time-consuming and cost-intensive measurements for plastically deformable foams.
Within the Helmholtz AI project FoAIm, we propose a method for predicting the mechanical properties of polyurethane foam using machine learning ultimately aiming to identify structure-property linkages. The method is based on a data set of experimentally validated simulations of reconstructed and algorithmically generated digital twins of polyurethane foam structures, which resolve microstructures in the micrometre range. ... mehrIn order to create an information-rich data base suitable for machine learning, the microstructures of the generated foams are systematically varied.
The data set is used to train a supervised machine learning algorithm to predict the mechanical properties of a foam from its microstructure. In order to ensure the plausibility of the predicted mechanical properties, explainable AI methods are applied to the trained model. These methods identify the most relevant regions within the microstructure for a given prediction, allowing to validate the plausibility of the trained network by comparing the basis of the prediction with known correlations between mechanical and geometric properties.
Through the application of explainable AI methods, the proposed strategy further reveals structure-property relationships within the foam that control the mechanical properties of the foam. This knowledege can be used to tailor the properties of the foam and to accelerate the development of novel materials for specific requirements through digital methods.