Machine Learning-Assisted Model Blending for Generalizable Turbulence Corrections

We present a machine learning-assisted model blending strategy for turbulence modeling, aimed at enhancing the generalizability of data-driven RANS corrections. Instead of applying a single correction model, our approach combines multiple specialized models ("experts") using spatially varying blending weights. These weights depend on local, non-dimensional flow features and are computed via two learning strategies: an a posteriori method using random-forest regression to predict Gaussian weights, and an a priori method based on soft memberships from Gaussian Mixture Model clustering and Xgboost. This enables smooth, adaptive transitions between models, allowing the final blended correction to respond to diverse turbulent flow regimes with improved accuracy and consistency.