Todd E. Combs

Date of Award


Document Type


Degree Name

Master of Science


Department of Operational Sciences

First Advisor

Jack M. Kloeber, Jr., PhD


This thesis describes the development of a methodology to minimize the cost of vitrifying nuclear waste. Pacific Northwest Laboratory (PNL) regression models are used as baseline equations for modeling glass properties such as viscosity, electrical conductivity, and two types of durability. Revised PNL regression models are developed that eliminate insignificant variables from the original models. The Revised PNL regression model for electrical conductivity is shown to better predict electrical conductivity than the original PNL regression model. Neural networks are developed for viscosity and the two types of durability, PCT-B and MCC-1 B. The neural network models are shown to outperform every PNL and Revised PNL regression model in terms of predicting property values for viscosity, PCT-B, and MCC-1 B. The combined Neural Network/Revised PNL 2nd order electrical conductivity models are shown to be the best classifiers of nuclear waste glass, i.e. they have the highest probability of classifying a vitrified waste form as glass when it actually did produce glass in the laboratory. Finally, five nonlinear programs are developed with constraints containing: (1) the PNL original 1st order models, (2) the PNL original 2nd order models, (3) the Revised PNL 1st order models, (4) the Revised PNL 2nd order models, and (5) the Neural Network/Revised PNL 2nd order electrical conductivity models. The Neural Network/Revised PNL 2nd order electrical conductivity nonlinear program is shown to minimize the total expected cost of vitrifying nuclear waste glass. This nonlinear program allows DOE to minimize its risk and cost of high-level nuclear waste vitrification.

AFIT Designator


DTIC Accession Number