Training of a physics-based thermo-viscoplasticity model on big data for polypropylene

Abstract

Research on data-driven constitutive models has demonstrated their outstanding ability to provide highly accurate predictions of the general stress-strain response after learning from data only. Here, we demonstrate that physics-based models can equally benefit from training procedures relying on big data. Specifically, we employ the thermo-mechanically coupled viscoplasticity model [Anand, L., Ames, N.M., Srivastava, V., Chester, S., 2009. A thermo-mechanically coupled theory for large deformations of amorphous polymers. Part 1: Formulation. International Journal of Plasticity] to describe the large deformation response of polypropylene. It combines both mechanism-based evolution equations and high mathematical flexibility. More than 100 constant velocity and strain rate jump experiments are performed on flat tensile specimens extracted from 3 mm thick isotactic polypropylene sheets. The exact cross-sectional area is measured with surround DIC, while an IR camera monitored the surface temperature field. The experiments typically reached true strains greater than 0.8 and cover temperatures and strain rates ranging from 25 to 85 °C and 10–4 to 100s-1, respectively. Training over 100,000 unique random combinations of experiments is performed to identify all model parameters. The effect of the training (and testing) subsets size and composition is carefully analyzed to ensure a high generalization ability. It is found that training based on 26 randomly-selected experiments leads to the most robust parameter estimates. The obtained model performs remarkably well on all our experiments (among which 70 % are unseen during training) with a root mean square error of less than 1.5 MPa. As a byproduct, we also found that there exists a subset of two specific experiments for training that lead to an equally accurate model for polypropylene.