Convolutional Neural Network Model for the Prediction of Plenum Temperature in a Waste Glass Melter

Abstract

Abstract Legacy radioactive tank waste is slated to undergo vitrification at the Waste Treatment and Immobilization Plate (WTP) using Joule-heated, ceramic-lined melters. A high-fidelity, computational fluid dynamics (CFD) model of the pilot-scale DM1200 melter has been developed to provide an understanding of the heat transfer and fluid dynamics within the WTP melters. Monitoring of the non-radioactive pilot-scale system has been primarily done through visual observations. However, visual observations will not be possible in the full-scale radioactive melter and process control will be based upon the measured plenum temperatures. Using the CFD model, the effect of the cold cap coverage on the plenum temperature can be assessed. The plenum temperature within the DM1200 is primarily driven by the amount and distribution of the cold cap coverage on the melt pool, since thermal radiation is the dominant mode of heat transfer at these temperatures. Plenum temperatures in the DM1200 obtained during pilot-scale testing by the Vitreous State Laboratory were used for model validation. A standard LeNet-1 convolutional neural network (CNN) is used to predict the spatial cold cap coverage of the model from the computed plenum temperature distributions derived from known cold cap topologies. With a 16 cm × 16 cm filter applied, an accuracy of 89% was achieved.