Characterisation of the spherulitic microstructure of semi-crystalline thermoplastics

Abstract Microstructure, as the connection between process and material properties in semi-crystalline thermoplastics, plays a crucial role in fabricating high-quality prats. Hence, it is essential to consider the microstructure by predicting the component properties. An integrative multi-scale simulation chain on an ICME platform was previously developed and successfully applied to predict the local, microstructure-dependant, effective elastic and thermal properties of an injection moulded plate from the isotactic polypropylene over its cross-section. However, the calculation time of several hours for small volume elements at the microscale seems to be an obstacle to employ it in real cases. Data-driven models, such as artificial neural networks are meant to be employed in the presented paper to solve this problem. The focus here is on the spherulitic microstructure of semi-crystalline microstructure predicted by the simulation in-house-developed software SphaeroSim as the link between solidification simulation and homogenisation. To build data-driven models parallel to these two simulation tools, the microstructure must be characterised, and meaningful features need to be extracted from it. This work aims to present general features for the characterisation of a spherulitic microstructure and build a data-driven model parallel to the solidification simulation, predicting the proposed features by a given cooling temperature history. A set of two-dimensional (surface area, circumference, Feret diameter, long Feret diameter, form factor, convex circumference, convexity, equivalent diameter and form deviation) and three-dimensional (volume, surface area, aspect ratio and sphericity) was calculated for a data set of about 500 simulations. Their distribution suggests that almost all of the spherulite have an ideally spherical shape. Therefore, some of the features can be replaced with each other. The achieved accuracies for each data-driven model relating to a predicted feature shows the great potential for substituting the original simulation tools for data-driven models, to immensely decrease the prediction time while keeping the high prediction accuracy.