Quantitative simulation is probably the major tool of industrial R&D studies. Computer codes are widely used to predict the behavior and the reliability of a complex system in given operating conditions or, in the design phase, to ensure that it will fulfill given required performances. Whatever their complexity, quantitative models are essentially physics-based and, consequently, deterministic. On the other hand, their inputs and/or the code itself may be affected by uncertainties of various nature which affect the final results and must be properly taken into account. Uncertainty analysis has gained more and more importance in industrial practice in the last years and is at the heart of several working groups and funded projects involving industrial and academic researchers.

We present hereby how the common industrial approach to uncertainty analysis can be formulated in a full Bayesian and decisional setting. First, we will introduce the methodological approach to industrial uncertainty analysis. On the other hand, we will recall some features of the Bayesian setting which prove useful in the industrial practice. We will particularly insist on the role of decision theory and we will show some usual misunderstandings and shortcuts when decisional issues are not formally stated.

These concepts will be shown by detailing the stakes and the results of some industrial examples.

Keywords: Uncertainty analysis; Bayesian inference; Decision theory; Heuristic predictive estimation

Biography: After having completed a PhD in statistical modelling applied to brain imaging, Dr. Merlin Keller is currently working as a research engineer in the industrial risk management department of EDF R&D. His main research interests are the use of Bayesian inference for industrial risk assessment, and Gaussian process emulation of complex computer codes. He is also involved in the development of the openTURNS open-source software package for statistical uncertainty analysis.