Measuring Predictive Capability of Computational Models: Foam Degradation Case Study

摘要

Statistical methods for evaluating the predictive capability of computational models are tested and illustrated for a Sandia National Laboratories case study pertaining to the degradation of polyurethane foam in a thermal environment. A newly developed computational model of this phenomenon is compared to a suite of nine experiments. The statistical analysis focuses on characterizing prediction-error as a function of experimental variables, primarily temperature. It is found that both predicted degradation-front velocity and the experimental data exhibit an approximate Arrhenius relationship, but with different slopes ("activation energies"). Statistical prediction intervals are obtained in each case and compared. The need for additional experimentation in order to resolve ambiguities is also discussed.