A Grid-Induced and Physics-Informed Machine Learning CFD Framework for Turbulent Flows

High fidelity computational fluid dynamics (CFD) is increasingly being used to enable deeper understanding of turbulence or to aid in the design of practical engineering systems. While such CFD approaches can predict complex turbulence phenomena, the computational grid often needs to be sufficiently refined to accurately capture the flow, especially at high Reynolds number. As a result, the computational cost of the CFD can become very high. It therefore becomes impractical to adopt such simulations for parametric investigations. To mitigate this, we propose a framework where coarse grid simulations can be used to predict the fine grid results through machine learning. Coarsening the computational grid increases the grid-induced error and affects the prediction of turbulence. This requires an approach that can generate a data-driven surrogate model capable of predicting the local error distribution and correcting for the turbulence quantities. The proposed framework is tested using a turbulent bluff-body flow in an enclosed duct. We first highlight the flow field differences between the fine grid and coarse grid simulations. We then consider a set of scenarios to investigate the capability of the surrogate model to interpolate and extrapolate outside the training data range. The impact of operating conditions and grid sizes are considered. A Random Forest regression algorithm is used to construct the surrogate model. Two different sets of input features are investigated. The first only takes into account the grid-induced error and local flow properties. The second supplements the first using additional variables that serve to capture and generalise turbulence. The global and localised errors for the predictions are quantified. We show that the second set of input features is better at correcting for the biases due to insufficient resolution and spurious flow behaviour, providing more accurate and consistent predictions. The proposed method has proven to be capable of correcting the coarse-grid results and obtaining reasonable predictions for new, unseen cases. Based on the investigated cases, we found this method maximises the benefit of the available data and shows potential for a good predictive capability.