Predicting mechanical properties of Foam structures using Machine Learning

Abstract

The use of composite materials such as a polyurethane aluminium sandwich structure is a promising method for weight reduction in novel vehicle concepts. Dimensioning such composite components, however, requires knowledge of the mechanical properties of the foam used. These properties are usually determined with the help of different mechanical testing methods. In order to replace these time- and cost-intensive experiments, the project presented here aims to develop a machine learning (ML) approach that identifies structure-property linkages in foams and can thus predict their mechanical properties based on the microstructure. To generate a suitable database for the training of an ML-algorithm, both experimental investigations of different foam structures are carried out and computational methods are applied. Via the experiments the mechanical properties of the foam structures are determined by means of tensile and compression tests while computer tomographic (CT) measurements are used to obtain high resolution images of the used foam samples. The resulting CT-scans are converted into digital representations of the microstructures and mechanical simulations as well as image analysis algorithms are applied using the PACE3D [1] simulation framework. In this way the used simulation model is validated. Additionally, new insights into the morphology of the foam structure are gained by extracting structure parameters such as the mean pore size, wall thickness and porosity. Further 3D foam structures are generated algorithm-based with defined structure parameters and their mechanical properties are computationally determined. The generated structures and their corresponding mechanical properties serve as the basis for training a suitable machine learning algorithm. Its implementation is realised within the framework of the research data infrastructure Kadi4Mat [2], which enables the structured storage of the created data as well as the implementation of the machine learning algorithm through automatable workflows. Kadi4Mat further enables the exchange of results and their comprehensible documentation through the use of metadata and ontologies. To enable both the reproducibility of the results and the transfer of the developed methodologies to other applications, all processes developed in the project are additionally archived in Kadi4Mat in the form of automatable workflows.