Hanford waste blending and the value of research: Stochastic optimization as a policy tool

A new approach to stochastic, combinatorial optimization is presented through its application to a contemporary policy problem: cleanup of radioactive wastes stored underground at the US Government's Hanford, WA, nuclear fuels processing site. Current plans call for the tank contents to be selectively combined prior to their immobilization in glass; such blending of wastes reduces the amount of extra material required for vitrification, therefore, decreasing the costs of processing and disposal. Uncertainty in the tank contents, the error inherent in the glass property models governing vitrification, and the computationally intensive nature of the problem, however, render determination of an optimal tank-blend assignment a challenge to existing optimization techniques. Previous studies have focused exclusively on minimization of processing and disposal costs, ignoring such policy-related dimensions as the value of reducing select sources of uncertainty. In addition, the stochastic framework employed by these studies could not guarantee that the glass property requirements (in the guise of model constraints) were met on more than a probabilistic basis. This paper presents a novel stochastic annealing–nonlinear programming framework that incorporates variance – a proxy for the opportunity costs of reducing uncertainty – as an attribute in its objective function. Compared with conventional mathematical programming algorithms, the new optimization framework is seen to be more robust, flexible, and efficient. The algorithm also facilitates analysis of the trade-off between minimizing processing and disposal costs, and reducing the expenses of achieving this savings. Specifically, the prediction error of the glass property models is found to be a more significant source of uncertainty than variation in tank component mass fraction estimates, and constraint violations are traced to specific requirements of the glass property models.