Department of Energy
University Information
Collaborators
Predictive Engineering and Computational Sciences
201 E. 24th Street
Austin, TX, 78712
Phone: 512-471-3168
Fax: 512-232-8067
Site Administrator
201 E. 24th Street
Austin, TX, 78712
Phone: 512-471-3168
Fax: 512-232-8067
Site Administrator
Algorithms Research
Activities
Modeling Domains
In current sensitivity studies, hundreds of forward solutions of the coupled multiphysics atmospheric entry problem are required to obtain even decoupled parameter sensitivity information. While such simple sensitivity studies are valuable, the straight-forward methods employed do not scale to solving more sophisticated uncertainty quantification problems.
To address the computational challenges inherent in calibration, validation and uncertainty quantification in models of complex systems, we are developing adjoint-based algorithms to accelerate uncertainty computations. These include algorithms to solve statistical inverse problems to obtain estimates of uncertain model parameters from experimental data, and uncertainty propagation problems to determine uncertainties of quantities of interest from uncertainties in models and input parameters.
These adjoint-based algorithms make use of the sensitivity of the quantities of interest to changes in the inputs. Such sensitivities are most easily computed by solving a linear adjoint equation to obtain an influence function that can then be used to rapidly calculate a complete local parameter sensitivity gradient. Once available, adjoint solutions can also be used for other purposes, such as a posteriori numerical error estimates, and goal-oriented grid refinement. All these capabilities are being implemented in a physics-independent, modular component of the libMesh software library for solving problems governed by partial differential equations.