Uncertainty quantification for data-driven turbulence modelling with Mondrian forests

Data-driven turbulence modelling approaches are gaining increasing interest from the CFD community. However, introduction of a machine learning (ML) model introduces new source uncertainty, ML itself. Quantification this uncertainty is essential since predictive capability data-driven diminishes when predicting physics not seen during training. In work, we explore suitability Mondrian forests (MF's) for modelling. MF's claimed to possess many advantages commonly used random forest (RF) algorithm, whilst offering principled estimates. An example test case constructed, with anisotropy constant derived high fidelity resolving simulations. Shapley values, borrowed game theory, interpret MF predictions. Predictive found be large in regions where training data representative. Additionally, exhibit stronger correlation errors compared an priori statistical distance measure, which indicates it better measure prediction confidence. The also calibrated and less computationally costly than estimated applying jackknifing Finally, predict Reynolds discrepancies convergent-divergent channel, subsequently propagated through modified solver. resulting flowfield predictions close agreement data. A procedure sampling forests' uncertainties introduced. Propagating these samples enables quantification output quantities interest.