Relation of validation experiments to applications

Model validation efforts often use a suite of experiments to provide data to test a model for predictive use for a targeted application. This suite may include individual experiments that provide only partial coverage of the physics relevant for a target application, experiments that are at different environmental conditions, and experiments that use idealized geometries to represent the target application.;Determining how well the experiments test the target application model is an issue that should be considered in designing such experimental suites. This paper provides methodology to partially address this issue. The methodology uses computational models for both the individual test suite experiments and a computational model for the target application, to assess coverage of the application by potential experiments.;The methods and procedures presented focus on three techniques to relate the test suite experiments to the target application. All techniques include metrics to quantify the tradeoff between the capability of validation experiments to resolve target application model predictions, and at the same time account for sensitivity of the results to model parameter uncertainty. The first technique uses validation model sensitivity matrices, application model sensitivity matrices, and matrix singular value decomposition, including the Moore-Penrose matrix pseudo-inverse. In addition, a limited investigation using non-dimensional sensitivity matrices is performed. The second technique uses the same matrices and matrix operations, but replaces the singular value decomposition approach of the first technique with an optimization using an analytically differentiable objective function. The third technique is an evolution of the second technique, with error metrics included in a non-linear, non-analytically differentiable objective function. Several numerical optimization methods to minimize the resulting objective functions are investigated.;Findings include comparisons of results when using sensitivity matrices composed of strictly first order sensitivity analysis matrices with results using second order sensitivity analysis matrices. Using first order sensitivity for subsequent analyses, all three techniques are compared by investigating the tradeoff between the ability of the suite of experiments, as represented by their models, to resolve the application model and the sensitivity of the results to model parameter uncertainty. Finally, this work explores refinement of sensitivity matrix composition using non-dimensional sensitivity matrices to represent a non-dimensional approach.