WP5 Algorithmic extensions of MLMC (UQ)
Objectives
The work package will focus on the theoretic developments needed to enable using the MLMC method together with unstructured meshes and time-dependent problems. Work will be done to make feasible the use of the technology in conjunction with LES-type technologies for the CFD solution. The use of U-RANS type approaches will be considered as a backup option in case the first option does not deliver on the expectations.
Theoretical work will also be done to take advantage of “deterministic” Adjoint-based error estimators (to be developed in WP2) in constructing their stochastic counterparts. Novel bleeding edge-approaches will also be deployed according to the expertise of the team leader.
Tasks
Task 5.1: Definition of basic interface for UQ (targeted to MLMC)
The first task of the WP will focus on the design of an interface between the different solvers and the scheduling engine. Such interface defines the “physics-based” point of view and will be the basis of the API development to be done in WP4.
Task 5.2: Construction of a python-driven MLMC engine, targeted to UQ
A python-driven engine for MLMC will be defined, probably based on previous software developments at EPFL which are soon to be open sourced. The engine will communicate to the CFD solvers through the API developed in WP4, so to take full advantage of the programming models defined by COMPSSs and HypeLoom in achieving an optimal scheduling of the tasks.
Task 5.3: Development of unstructured MLMC capabilities for Uncertainty Quantification
The task will focus on Uncertainty Quantification targeting simplified CFD-like problems, so to minimize run time. This simplification will include (for example) employing steady state solvers and/or Potential-flow or Euler formulations. The goal is to test the novel developments including the use of the adjoint-driven mesh adaptation. This task will focus mainly on shape-related uncertainties.
Task 5.4: MLMC for time dependent problems
This task will expand the capability developed in the previous task to the simulation of time dependent problems. This will imply interacting with the developments in WP3 and considering a mesh variability within the simulation. Simplified CFD solvers will still be employed at this stage as long as this is compatible with the physical problem being considered. This will also put to test the cost-models to be developed for the CFD solvers.
Task 5.5: Uncertainty Quantification in wind engineering
The task focuses on tailoring the UQ framework for the application test case chosen as showcase (wind engineering). It will imply considering both the statistic variability in the wind and taking into account uncertainties in the geometry definition. Full-precision Navier-Stokes solvers will be needed in this context, thus implying that runs at full-machine scale (as available) will be targeted.
