Simulation of the Heat Transfer during the Casting Process by Mirror U-Net Models

Abstract

Deep learning has achieved great progress in image recognition, segmentation, semantic recognition, and game theory. It also shows potential to solve scientific computing such as simulation problems in engineering. On the other hand, the numerical simulation method requires constitutive modelling, involves a huge computation volume and takes a long time. In this paper, two mirror U-Net models were proposed for the simulation of the heat transfer during the casting process. These models include an upper U-Net branch for the treatment of the geometries of casting, mold, and chill, and a lower U-Net branch for the treatment of the initial temperature field. Their difference is whether the bottoms of upper and lower U-Nets are shared. These two branches tackle the problems involving the input of a geometrical model which consists of three types of materials and the input of an initial or current temperature field image. These models were trained and validated with a big database with hundreds of casting shapes. The prediction results show that the average accuracy reaches 98.8%.