Tuning of data-driven algebraic turbulence models to improve RANS prediction in separated flows

The present thesis aims at the development of data-driven Explicit Algebraic Reynolds Stress Models (EARSM) to improve the prediction of the Reynolds stress tensor in separated flows, with particular attention to separation-induced transition phenomena relevant to turbomachinery applications. A comprehensive experimental database, consisting of 87 laminar separation bubbles measured through Time-Resolved Particle Image Velocimetry (TR-PIV) at the University of Genoa, was first employed for model training. Sparse Bayesian Learning (SBL) was used to identify parsimonious models able to reproduce the anisotropy properties of the Reynolds stress tensor based on both classical invariants and extended flow-feature libraries. Starting from a baseline model with 37 predictors, reduced models containing 13 and 15 predictors were obtained for long and short separation bubbles, respectively. The inclusion of flow features among the predictor library will show improved accuracy compared to invariant-based formulations. To enhance model generalizability across different flow regimes, a space-dependent Model Aggregation (X-MA) [1] strategy was introduced. This method employs SBL not only for turbulence model calibration but also for determining spatially varying weights that combine the predictions of individually trained models. Initially, the technique was applied to weight the solutions of a model specific for long bubbles and one for short bubbles. The goal was to extend the generalizability of the aggregated model to long and short bubbles as well as to bubbles exhibiting intermediate behavior between the two. The aggregation procedure was further extended by combining the predictive capabilities of two models trained on two different datasets: on the experimental flat-plate dataset without a distinction in bubble size and on a second numerical LES dataset. The latter included 15 separation-bubble conditions evolving over a flat plate with a rounded- nose leading edge. In the first case, separation is mainly driven by the pressure gradient, whereas in the second case it is instead related to the geometry of the plate leading edge. The aggregated model exhibited improved predictive capabilities when applied to a third unseen flow configuration that did not participate to the training. Finally, an algebraic model for intermittency was calibrated using the experimental laminar separation bubble dataset. When coupled with the trained turbulence model, this intermittency model dampens low-frequency oscillations in the upstream part of the separated shear layer, ensuring that non-zero Reynolds stress values arise only where true turbulence is present. Overall, the outcomes of this work demonstrate the potential of ML-driven turbulence modelling to bridge the gap between high-fidelity simulations and practical RANS solvers. The developed models contribute to a more accurate and efficient prediction of separated and transitional flows, paving the way for next-generation data-informed tools in aerodynamic design and sustainable propulsion system development.