Solidification Estimation using Spherical Neural Networks
Monday, June 19, 2023, 10:20 AM - 10:40 AM
Max Bell Theatre
Maximilian Erber Constantin Bauer

Due to their high degrees of freedom castings are predestined to apply structural optimization during their design process. Until today the integration of process simulations in structural optimization is limited due to high computational cost and therefore often neglected in the beginning design process. This leads to the need of surrogate models, which allow a fast and simplified evaluation of design proposals during the optimization in order to improve the integration. In this article, a novel approach is introduced that estimates the solidification time of randomly created geometries solely based on the casting geometry. The approach uses ray-tracing methods to calculate the distance function along preset directions. The estimated solidification time is calculated using an Icosahedral Convolutional Neural Network (CNN). The training data is obtained by several thousand solidification simulations using the optimization toolkit of a commercial casting simulation software combined with further data augmentation. The model is experimentally validated for five different geometries in the sand casting process.

Moderated by: Andre Phillion / Matt Krane