15:45
MS30: Numerical methods for PDE-constrained and controlled optimization problems with appplications (Part 3)
Chair: Maya Neytcheva
15:45
25 mins
|
Hyper-differential sensitivity analysis for PDE-constrained optimization
Bart vanBloemenWaanders, Joey Hart
Abstract: Hyper-differential sensitivities (HDS) analyze the dependence of PDE-constrained optimization solutions to parameter perturbations. Such analysis may be used to prioritize uncertainties in the service of data acquisition, uncertainty quantification, and model development. Low rank structure is exploited through a Singular Value Decomposition which is numerically implemented with randomized algorithms using multi-level parallelism in C++. HDS is demonstrated (1) to prioritize the influence of uncertain boundary conditions and material properties on control strategies, (2) to analyze the stability of optimal solutions under uncertainty, and (3) to augment optimal experimental design for data acquisition.
|
16:10
25 mins
|
Reduced Order Modeling for Nonlinear PDE-constrained Optimization using Neural Networks
Nikolaj Takata Mücke, Lasse Hjuler Christiansen, Allan Peter Karup-Engsig, John Bagterp Jørgensen
Abstract: With simulation based decision making playing an increasingly important role in science and engineering the demand for fast and reliable computational schemes is increasing. This is e.g. the case for nonlinear model predictive control (NMPC) where real-time multi-query solutions are essential. However, in cases where the mathematical model is of high dimension in the solution space, e.g. for solution of partial differential equations, black-box optimizers are rarely sufficient to get the required online computational speed.
In this talk, I will present a reduced order modeling approach, based on proper orthogonal decomposition (POD) and artificial neural networks (ANN), to address before mentioned problems associated with nonlinear PDE-constrained optimization. The role of POD is to identify a lower dimensional representation of the solution manifold while the ANN is used for approximating a parametrization of the low dimensional manifold. Thus, leading to an equation free online model. I will consider a specific nonlinear time dependent PDE-constrained optimization problem and assess the performance and potential of the proposed strategy.
|
16:35
25 mins
|
PDE-Constrained Optimization: Optimal control with L_1-regularization, state and control box constraints
Ivo Dravins, Maya Neytcheva
Abstract: We present a method for solving optimal control problems constrained by a partial differential equation, where we simultaneously impose sparsity-promoting $L_1$ regularization on the control as well as box constraints on both the control and the state. We focus on numerical implementation aspects and on preconditioners used when solving the arising linear systems.
|
17:00
25 mins
|
An alternative method to impose state and control constraints in PDE-constained optimization problems
Maya Neytcheva
Abstract: We compare two approaches to impose additional constraints on the state and the control variable in the framework of PDE-constrained optimization problems.
|
|
