Prediction of melter cold-cap topology from plenum temperatures with computational fluid dynamics and machine learning

Abstract

A computational fluid dynamics (CFD) model of a pilot-scale waste-glass melter was used to generate input data for several different machine-learning models to predict the cold-cap coverage from plenum temperatures. This methodology could serve as useful to provide nonvisual feedback for operational control. The machine-learning models tested include an artificial neural network (ANN), a convolutional neural network (CNN), a random forest (RF) algorithm, and a support vector machine (SVM) method. CFD simulations with randomized cold-cap coverage were run to generate plenum-temperature distributions. The topology of the cold cap was predicted with each method using an input layer of selected temperature locations. The ANN was used to predict the cold-cap coverage percentage with an accuracy of 1.2%. The accuracy of the various machine-learning algorithms was dependent on the filter resolution employed. When using a low resolution (16-cm filter), the CNN and ANN methods produced the best accuracy, with errors of 6.98% and 6.58%, respectively. At finer levels of resolution (4-cm filter), the RF method produced the best accuracy, with an error of 8.12%. The ANN and SVM methods required less computational time to train than either the CNN or RF methods.