Towards Local Application of Data-Driven Turbulence Modeling based on Field Inversion and Machine Learning

Kurzfassung

Recent advancement in data-driven turbulence modeling has shown the potential of the usage of machine learning techniques to enhance classical Reynolds-averaged Navier-Stokes (RANS) turbulence models. However, one of the main limitations of the current data-driven approaches includes their limited applicability to flows that are very similar to the training cases. Moreover, when dataaugmented model is applied to out-of-training flow scenarios, the predictive accuracy is often harmed, i.e., the data-driven turbulence model provides worse prediction than the baseline model. While efforts have recently been made to address such robustness and generalizability issues of data-driven methods by improving the training strategy (e.g., [1,2]), we present a sensor-based approach towards more general data-driven models here, namely the local activation or deactivation of (different) model augmentations depending on the local flow state. For this, physics-based classical sensors and also machine learning classifiers could potentially be used.