Prediction of long-term creep modulus of thermoplastics using brief tests and interpretable machine learning

Abstract

The prediction of creep behavior plays a critical role in the design of thermoplastic materials intended for prolonged use. The creep modulus, which describes the relationship between stress and strain that a material experiences over time, is a key property to determine the long-term thermo-mechanical performance of thermoplastics. Due to the time-consuming and resource-intensive nature of testing for this property, the present work investigates the potential of data-driven techniques as an alternative approach. To accomplish this, a dataset comprising more than 400 distinct thermoplastic grades was obtained from CAMPUS® online open database. Then, various interpretable machine learning models (linear regression, decision trees, random forests, XGBoost, and LightGBM) were evaluated to predict the long-term creep modulus with data from brief tests. To accurately assess the models’ ability to generalize to new data, rigorous model evaluation techniques such as cross-validation and group-splitting were employed, showing that various algorithms can predict the creep modulus with scores above 0.99. Interestingly, linear regression not only matches but, in some cases, also surpasses the performance of more complex models, while being the most simple and interpretable. The present work demonstrates that machine learning can bypass the most lengthy creep tests; reducing costs, energy consumption, material waste, and product development time.