Explainable Prediction of Mechanical Properties in Porous Microstructures

In automotive design high-performance sandwich composite materials that combine light weight and high energy dissipation play an important role. In particular, sandwich composites with a polyurethane foam core and a metallic face material exhibit excellent performance. To dimension such composites, it is necessary to know the mechanical properties of the constituent materials. The determination of the properties of the polyurethane foam core, however, commonly requires extensive investigations as they greatly depend on the foam’s microstructure that varies significantly with manufacturing equipment and process parameters. Experimental testing can determine the mechanical properties, but offers a time-consuming and cost-intensive remedy to the underlying problem.

To avoid these resource intensive investigations, we propose a deep-learning strategy that can efficiently predict the mechanical properties of foams solely from their microstructures.
The basis for this approach is an algorithmically generated lookup table with digital twins of foam variants in the form of 3D models that resolve the microstructures on a micrometre scale. ... mehrExplainable AI methods applied to the database establish a comprehensible connection between the predicted mechanical properties and the microstructure domains of most relevance for a specific prediction. Thus, we validate the trained network in terms of its plausibility by comparing the database of the prediction to known correlations between mechanical and geometric properties. Ultimately, the proposed strategy yields structure-property linkages of foams, which control mechanical properties and can be used to tailor foam properties to specific requirements.