Neural network model describing the temperature- and rate-dependent stress-strain response of polypropylene

Abstract

A machine learning based model is proposed to describe the temperature and strain rate dependent response of polypropylene. A hybrid modeling approach is taken by combining mechanism-based and data-based modeling. The “big data” required for machine learning is generated using a custom-made robot-assisted testing system. Numerous large deformation experiments are performed on mildly-notched tensile specimens for temperatures ranging from 20 to 80 °C, and strain rates ranging from 10−3 to 10−1/s. Without making any a priori assumptions on the specific mathematical form, the function relating the stress to the viscous strain, the viscous strain rate and temperature is identified using machine learning. In particular, a back propagation algorithm with Bayesian regularization is employed to identify a suitable neural network function based on the results from more than 40 experiments. The neural network model is employed in series with a temperature-dependent spring to describe the stress-strain response of polypropylene. The resulting constitutive equations are solved numerically to demonstrate that the identified model is capable to predict the experimentally-observed stress-strain response for strains of up to 0.6.