Predicting Ductile–Brittle transition temperatures for polyolefins using convolutional neural networks and instrumented notched Charpy experiments

For polyolefins, a transition from ductile- to brittle fracture behavior can be observed at decreasing temperatures, which is typically characterized by the Ductile-Brittle Transition Temperature (DBTT). For each material, the DBTT needs to be determined by performing multiple impact tests across wide temperature ranges. In this study, we utilize Convolutional Neural Networks (CNN) to predict the DBTT from single Instrumented Notched Charpy (INC) experiments instead. Our dataset consists of over 6732 INCs encompassing 132 Polyethylene (PE) and Polypropylene (PP) compounds. For each of these compounds, we examine 51 impact tests at temperatures ranging from -40 °C to 60 °C. We train models on combinations of instrumentation signals and fracture surface images. We test the trained networks on 1683 INCs, where we achieve a R2-score of 0.96 and an average error below 5 °C of predictions compared to experimentally determined DBTTs. The presented methodology provides a flexible way of predicting DBTTs in case of non-instrumented testing or if thermal specimen conditioning is not possible. This allows for fast screening of new materials.