Traditionally, simulating turbulent flow requires extensive computational resources to produce high-quality results using Direct Numerical Simulation (DNS) or Large Eddy Simulation (LES). While these methods yield highly accurate predictions for quantities such as the velocity of the flow, they are computationally infeasible for industrial applications. Instead, lower-fidelity Reynolds Averaged Navier-Stokes (RANS) models, such as k−ε, provide a compromise between accuracy and speed. However, it is well understood that these RANS models are inaccurate on flows with separation, curvature, and large pressure and buoyancy gradients [8].

Therefore, the goal of this thesis is to develop a proof-of-concept RANS turbulence model closure that would overcome the shortcomings of traditional turbulence models. Instead of using an experimental approach to inform the new turbulence model, a data-driven approach was adopted by using statistical techniques such as machine learning. The developed model proposed in this thesis is based on an Explicit Algebraic Stress (EASM) formulation of RANS, which calculates the anisotropic Reynolds stress.

The machine learning model is trained on LES data for both isothermal and stratified open channel flow. The proposed turbulence model uses a neural network, which underwent hyperparameter optimisation to form what is known as a Tensor Basis Neural Network (TBNN). This TBNN was evaluated based on three factors: accuracy, generalisability and interpretability. It is concluded that the TBNN made more accurate predictions for the simpler isothermal flow case, however, even the stratified predictions are said to be more accurate than the predictions which would be made using traditional RANS models.

Moreover, statistical inference shows the TBNN to be generalisable, with reliable average predictions for anisotropic Reynolds stress possible. Finally, the quality of the input variables is measured using Partial Dependence Plots (PDP). Here, it is found that the conditions for Galilean invariance are met and that a wall-based Reynolds number and a variable measuring the stratification effects are essential to making accurate predictions. While this study confirms that a data-driven turbulence model is viable, further variable selection and a posterior study to predict the fluid velocity is required.