Space-dependent turbulence model aggregation using machine learning

TL;DR

This paper introduces a machine learning-based approach to combine multiple turbulence models spatially, improving prediction accuracy and quantifying uncertainty in flow simulations, demonstrated on compressor cascade flows.

Contribution

A novel data-driven method using Random Forests to optimally weight turbulence models spatially, enhancing ensemble predictions and uncertainty estimation.

Findings

Aggregated turbulence model predictions outperform individual models.

Uncertainty estimates align well with high-fidelity data.

Method effectively adapts to unseen flow configurations.

Abstract

In this article, we propose a data-driven methodology for combining the solutions of a set of competing turbulence models. The individual model predictions are linearly combined for providing an ensemble solution accompanied by estimates of predictive uncertainty due to the turbulence model choice. First, for a set of training flow configurations we assign to component models high weights in the regions where they best perform, and vice versa, by introducing a measure of distance between high-fidelity data and individual model predictions. The model weights are then mapped into a space of features, representative of local flow physics, and regressed by a Random Forests (RF) algorithm. The RF regressor is finally employed to infer spatial distributions of the model weights for unseen configurations. Predictions of new cases are constructed as a convex linear combination of the underlying models solutions, while the between model variance provides information about regions of high model uncertainty. The method is demonstrated for a class of flows through the compressor cascade NACA65 V103 at Re~3e5. The results show that the aggregated solution outperforms the accuracy of individual models for the quantity used to inform the RF regressor, and performs well for other quantities well-correlated to the preceding one. The estimated uncertainty intervals are generally consistent with the target high-fidelity data. The present approach then represents a viable methodology for a more objective selection and combination of alternative turbulence models in configurations of interest for engineering practice