Towards an integrated machine-learning framework for model evaluation and uncertainty quantification

Abstract

We introduce a new paradigm for treating and exploiting simulation data, serving in parallel as an alternative workflow for model evaluation and uncertainty quantification. Instead of reporting simulations of base-case and specific variations scenarios, databases covering a wide spectrum of operational conditions are built by means of machine-learning using sophisticated mathematical algorithms. While the approach works for all sorts of computer-aided engineering applications, the present contribution addresses the CFD/CMFD sub-branch, with application to a widely used benchmark of convective flow boiling. In addition to comparing simulation and experimental results on a case-by-case basis, machine-learning is used to create their respective (CFD and experiment) data-driven models (DDM), which will in a later stage serve for assessing the predictive performance of the CFD models over a wider range of experimental conditions, hence providing a high-level classification of their range of applicability.