Polyolefin ductile-brittle transition temperature predictions by machine learning

Abstract

Polymers show a transition from ductile-to brittle fracture behavior at decreasing temperatures. Consequently, the material toughness has to be determined across wide temperature ranges in order to determine the Ductile-Brittle Transition Temperature This usually necessitates multiple impact experiments. We present a machine-learning methodology for the prediction of DBTTs from single Instrumented Puncture Tests Our dataset consists of 7,587 IPTs that comprise 181 Polyethylene and Polypropylene compounds. Based on a combination of feature engineering and Principal Component Analysis, relevant information of instrumentation signals is extracted. The transformed data is explored by unsupervised machine learning algorithms and is used as input for Random Forest Regressors to predict DBTTs. The proposed methodology allows for fast screening of new materials. Additionally, it offers estimations of DBTTs without thermal specimen conditioning. Considering only IPTs tested at room temperature, predictions on the test set hold an average error of 5.3°C when compared to the experimentally determined DBTTs.