This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The production of high-quality glass demands precise control over furnace parameters, raw material inputs, and a thorough assessment of the final glass product. Quality and defect levels are crucial factors influencing efficiency, cost, waste, and sustainability in glass manufacturing. Given the multitude of furnace parameters, often of the order 100, determining quality through numerical simulation alone remains a challenging problem. Adding to this complexity is the time delay between the formation of molten glass and the detection of defects in the final product. To achieve good quality glass, the residence time of the glass in the furnace needs to be at least 8-12 hours. This residence time is not constant and highly depends on the process parameters. Changes made to the inputs will affect the quality of the final product at a much later stage. A tool that predicts the upcoming glass quality as a function of current and previous inputs supports optimal furnace performance. CelSian addresses these challenges by integrating machine learning with furnace simulation to predict glass quality based on process parameters and user input variables. Delfos, powered by CelSian'™s CFD simulation package GTM-X, simulates furnace dynamics from user inputs, capturing changes in parametric values over time. By training an AI model on a large dataset, Delfos predicts defect counts over time. This approach combines the physics-driven simulation of GTM-X with the predictive power of AI, offering a novel pathway for proactive quality control in glass production. An important aspect of the development is a parallel, governmental-funded, project to speed up the CFD code significantly. This is done through AI-enforced solvers and usage of GPU'™s. The currently ongoing research projects for glass quality prediction show promising results. This presentation will show some of the results achieved and a forecast for future improvements. Also, the implementation of AI to speed up the CFD is briefly discussed.