Inverse Uncertainty Quantification for Sounding Rocket Dispersion

Sammanfattning

Sounding rocket impact points are subject to dispersion due to uncertainties in simulation model parameters and perturbations of the rocket trajectory during flight. Estimating the area of dispersion assumes that associated model uncertainties and magnitude of perturbations have already been inferred. In this thesis, a method to inversely quantify uncertainty in rocket simulation models based on launch data is presented. We take on a probabilistic approach based on Bayesian hierarchical modeling, to address both epistemic and aleatory uncertainty while incorporating prior knowledge about the modeled system. Bayesian computational techniques, including Markov Chain Monte Carlo simulations and modular Bayesian analysis, are accounted for and employed in numerical case studies. Surrogate deep neural network models are shown to ease otherwise infeasible computational burden that posterior distribution exploration suffers from. Numerical experiments are carried out based on actual launch data from Esrange Space Center, serving as validation of the methodology and providing posterior distributions of the target dispersion parameters. The results imply almost certainly that the currently used dispersion parameters can be reduced, for all considered sources of uncertainty in the study. Updating said parameters accordingly yields a potential 20% decrease in theoretically estimated dispersion area, which is in good agreement with empirical observations.