WP6 Optimization under uncertainties
WP6 focuses on “optimization under uncertainties” and as such, it embodies the ultimate and most ambitious goal of ExaQUte. The workpackage will be built on the top of the available deterministic optimization expertise, based mainly on the use of gradient-base optimization techniques.
The challenge of the WP will be to define a stochastic counterpart of the currently available deterministic optimizers. Developments will focus on simple objective functions, which will most likely allow to maximize the level of parallelism and thus permit taking advantage of the largest machines.
Task 6.1: Deterministic Optimization packages
The task will focus on adapting the current deterministic optimization capabilities to the MLMC API. The task includes making available the “sensitivities” wrt changes in the design variables once chosen a given, determistic, objective function.
Task 6.2: Computation of “Stochastic Gradients” to be used in performing optimization
The task focuses on the stochastic correspondent of the deterministic sensitivities (gradients), which will be required by the gradient optimizer and will rely on the MLMC engine developed in WP5. Alternatives will be considered on the base of the current state-of-the-art on the subject.
Task 6.3: Stochastic optimization for simplified problems
This task focuses on the first step towards the final optimization goal. It will tackle the stochastic optimization (Optimization Under Uncertainties) for simplified CFD-like problems. Steady state solution and simpler CFD models (potential and Euler flow) will be used at this stage to enable a lower computational effort while testing the infrastructure.
Task 6.4: Stochastic Optimization for unsteady problems
The task will consider unsteady configurations, taking into account shape-related uncertainties (or other sources as deemed necessary). Simplified flow models will still be considered as possible.
Task 6.5: Stochastic Optimization for wind engineering
The task is preparatory to the application, and will focus on tailoring the computational infrastructure of the specific case of Wind Engineering which represents the application test case of choice for the project. The use of full-scale machines and of full-precision solvers will be targeted here.