UQ at CERFACS
The activity on uncertainty quantification at CERFACS aims to estimate uncertainties in computational models requiring large computational resources. We consider applications related to environmental risk, data assimilation and optimization in a high performance computing context.
The principal actions are:
- Ensemble experimental design for calibration of uncertain model parameters
- Development cost-effective Surrogate Model methodologies for computationally intensive problems
- Sensitivity analysis, optimization and data assimilation
- Development of efficient algorithms for stochastic parameter estimation
These actions are implemented for the following applications:
- Surrogates for high dimensional problems in hydraulics and aerodynamics.
- Efficient, scalable and resilient domain decomposition algorithms for exascale computing
- Surrogate models for large-eddy simulations of pollution dispersion in the atmospheric boundary layer
- Developing Multifidelity-MLMC algorithms for industrial CFD, multidisciplinary systems and geosciences.
- Building simple climate models to explore data constraints on climate projections
Models are approximations of systems. Confidence in model simulations requires an understanding of their uncertainties.
Uncertainties are generally classified according to their nature. There are several sources of error:
- errors related to simplifications of equations (dimensional reduction, empirical parameterization)
- errors related to the numerical schemes
- errors related to spatio-temporal discretization.
Models must be supplemented by data and parameters that describe the system, initial conditions and its boundary conditions. The physical parameters govern the laws of the system – and sometimes, these data and parameters are only partially and approximately observed and known.
Errors can be further classified into two groups :
- Epistemic errors linked to a lack of knowledge of system processes
- Random errors linked to the stochastic nature of the system
Our aim is to understand and quantify the extent to which epistemic and random uncertainties affect the model response. At CERFACS, we develop a suite of tools to better understand this propagation of uncertainty.
- Multi-fidelity Monte-Carlo methods to allow the efficient estimation of probabilistic parameter distributions for industrial and geoscience applications
- Novel surrogate models to allow rapid exploration of parameter response in global climate and CFD problems.
- Risk assessment tools bridging scales for climate impacts, connecting global uncertainties to local impacts.
At CERFACS, these activity are transversal, with applications ranging from flood forecasting to wildland fire propagation, climate projection uncertainty and combustion chamber ignition models.