TY - JOUR

T1 - Artificial neural network model for estimating mold flux melting temperature

AU - Mapelli, Carlo

AU - Mombelli, Davide

AU - Dall'osto, Gianluca

AU - Cho, Jungwook

AU - Gruber, Nathalie

AU - Marschall, Irmtraud

AU - Cornille, Maïte

AU - Alloni, Marco

AU - Carli, Ricardo

PY - 2024/10/30

Y1 - 2024/10/30

N2 - The accuracy of the current models for the calculation of the melting temperature of the mold flux shows that there is still room for improvement, given that their accuracy could not be satisfactory enough to keep up with the current industrial needs. In this work the use of artificial neural networks for data prediction is explored. The network acts as a "black box" capable to predict the melting temperature determined by complex physical interaction among the involved chemical species composing the flux. The network is trained by learning from real experimental data provided by different research groups through hot stage microscopy. The data was tested first within its respective batches and then tested as a single aggregate data batch. After testing and optimization of the networks' parameters, an acceptable level of accuracy was reached because the estimated melting temperatures point out an average error lower than 30 K if compared to measured data. This opens the possibility for the development of a standalone application that can be used for reference. In order to open the possibility for further improvements of this study the paper shares and makes public the values contained in the matrixes connecting the nodes of neural networks.

AB - The accuracy of the current models for the calculation of the melting temperature of the mold flux shows that there is still room for improvement, given that their accuracy could not be satisfactory enough to keep up with the current industrial needs. In this work the use of artificial neural networks for data prediction is explored. The network acts as a "black box" capable to predict the melting temperature determined by complex physical interaction among the involved chemical species composing the flux. The network is trained by learning from real experimental data provided by different research groups through hot stage microscopy. The data was tested first within its respective batches and then tested as a single aggregate data batch. After testing and optimization of the networks' parameters, an acceptable level of accuracy was reached because the estimated melting temperatures point out an average error lower than 30 K if compared to measured data. This opens the possibility for the development of a standalone application that can be used for reference. In order to open the possibility for further improvements of this study the paper shares and makes public the values contained in the matrixes connecting the nodes of neural networks.

U2 - 10.2355/isijinternational.ISIJINT-2024-151

DO - 10.2355/isijinternational.ISIJINT-2024-151

M3 - Article

JO - ISIJ international

JF - ISIJ international

SN - 0915-1559

IS - Stand: 27. November ???

ER -