Machine learning-based prediction of glass color in an industrial furnace

Abstract

Color consistency is a critical quality attribute in industrial glassware manufacturing, yet predictive modeling of glass color, particularly the Commission Internationale de l’Éclairage (CIELab) a * and b * coordinates, is limited. This work presents a four-stage machine learning framework for industrial-scale precise prediction of glass color coordinates a * and b * values using glass furnace operational data. The methodology integrates systematic data preprocessing, including missing value handling, multivariate outlier detection, Shapley additive explanation (SHAP)-based feature importance analysis, and lag feature engineering, to capture temporal effects. A quadrant-based classification model is introduced to encode regional color behavior. Multiple tree-based and linear regression models are evaluated using manual and Bayesian hyperparameter optimization. Model performance is assessed using R 2, mean absolute error (MAE), mean squared error (MSE), and bias metrics, supported by a predicted versus actual analysis. This work is the first to systematically benchmark multioutput ML models for predicting CIELab color coordinates from industrial furnace operational data, demonstrating the feasibility and practical value of data-driven color prediction in continuous glass manufacturing.