PNNL

Abstract

Part 1 (Cooley and Piepel, 2003a) describes the first stage of a two-stage experimental design to support property-composition modeling for high-level waste (HLW) glass to be produced at the Hanford Site in Washington state. Each stage used a layered design having an outer layer, an inner layer, a center point, and some replicates. However, the design variables and constraints defining the layers of the experimental glass composition region (EGCR) were defined differently for the second stage than for the first. The first-stage initial design involved 15 components, all treated as mixture variables. The second-stage augmentation design involved 19 components, with 14 treated as mixture variables and 5 treated as non-mixture variables. For each second-stage layer, vertices were generated and optimal design software was used to select alternative subsets of vertices for the design and calculate design optimality measures. A model containing 29 partial quadratic mixture terms plus 5 linear terms for the non-mixture variables was the basis for the optimal design calculations. Predicted property values were plotted for the alternative subsets of second-stage vertices and the first-stage design points. Based on the optimality measures and the predicted property distributions, a ‘best’ subset of vertices was selected for each layer of the second-stage to augment the first-stage design.