A Data-Driven Approach for the Fast Prediction of Macrosegregation

Macrosegregation is of great importance to study due to its negative impact on the quality of casting. Although numerical models can predict macrosegregation during alloy solidification, solving the partial differential equations is rather time-consuming. Thus, numerical simulations are almost inoperable for the real-time online monitor-adjustment in industrial production, where the prediction is expected to be completed in an extremely short time. To overcome this challenge, a data-driven approach based on deep learning is proposed to predict the macrosegregation pattern under specific input parameter(s). Based on limited simulation results, this approach focuses on mining certain patterns within massive data, and thus enables fast predictions of macrosegregation, by incorporating a convolutional neural network autoencoder with a fully connected neural network. The best prediction accuracy is achieved after clarifying the effects of the error metric and the convolutional filter size. This method can predict the macrosegregation distribution in less than 0.1 second, and the accuracy is comparable to the conventional numerical simulations. The data-driven approach developed in this work shows instantaneity and adequate accuracy in the prediction of macrosegregation and could be a promising method for the application in the direct visualization and quality control of casting.