Gradient-enhanced surrogate models for efficient uncertainty quantification
Brian Lockwood, University of Wyoming
Uncertainty quantification is increasingly becoming an important avenue of research in the field of computational science. Uncertainty analysis provides valuable information which can greatly enhance the design and analysis capability of simulations by providing margins of error for crucial output quantities and a metric by which simulations can be adapted to improve the precision of the results. Typically, exhaustive sampling is used to quantify the uncertainty in simulation outputs based on uncertain input parameters. For high-fidelity simulations, this exhaustive sampling can be prohibitively expensive as it requires numerous evaluations of the computational model. In order to reduce the cost associated with exhaustive sampling, evaluations can be drawn from an inexpensive surrogate model. This surrogate model approximates the relationship between a simulation output and inputs based on a relatively limited number of simulation results. In order to enhance the accuracy of these surrogate models, derivatives of simulation outputs with respect to input parameters may be incorporated into the training of the model. With the use of adjoint-based approaches and automatic differentiation, these derivative values can be calculated at approximately the same cost as the original simulation, providing additional information about the output without significantly increasing sampling costs. For this poster, the use of gradient enhanced surrogate models for approximating the output of engineering simulations will be presented, with a particular emphasis on examples drawn from nuclear engineering and aerothermodynamics.
Abstract Author(s): Brian Lockwood, Mihai Anitescu, and Dimitri Mavriplis