Machine-Learning-Based Reduced Order Modeling for Operational Analysis of Industrial Glass Melting Furnaces Using CFD Solutions

Abstract

Computational Fluid Dynamics (CFD) models play a vital role in the design of industrial glass melting furnaces, offering insights into energy consumption, glass quality, temperature distribution, and refractory wear. However, the considerable computational expense associated with the large time and length scales involved in the glass melting process prevents practical utilization of those models in daily operation of the furnaces. This study presents a novel approach to address this challenge through the development of a machine-learning-based Reduced-Order Model (ROM) utilizing parametric data obtained from a CFD model of a glass melting tank of a furnace. Key operational parameters, namely pull rate, heat flux from combustion space, and electrical potential difference to supply electrical power, are chosen to create a CFD solution dataset, as they change the boundary conditions of the CFD model and, consequently, the field solution data. An autoencoder structure incorporating convolutional neural networks is established to learn and predict temperature and velocity field data. Then, the decoder section of the autoencoder is connected to the operational parameters through an auxiliary neural network. The performance of the reduced-order model is assessed for both interpolation and extrapolation using additional CFD solutions. Comparison between the field data generated by the ROM and the ground-truth CFD solutions indicates less than 1\% deviation, proving that the ROM’s capability to serve as an effective analysis tool for daily furnace operation. Furthermore, the ROM demonstrates significant advancements in solution time, up to third order, further enhancing its practical utility.