Hybrid Machine-Learning and Computational Fluid Dynamics Methods in the Energy Industry

Improving the modelling and simulation of turbulent and multiphase flows is a key issue within the energy industry due to their relevance and impact within the field. Three-phase flows involving air, oil and water are particularly common but are poorly understood due to their complexity. Gaining a better understanding of these flows is crucial in order to accurately predict the occurrence of specific multiphase flow regimes and increase efficiency in practical applications.
Most analysis relies on Computational Fluid Dynamics (CFD) modelling, however the large computational costs associated with this can often be impractical for industrial uses involving large number of hyperparameters. Consequently, there is a need for hybrid models which combine physics based CFD models with statistical methods in order to reduce computational costs and account for uncertainty. This can be achieved through the use of Machine Learning (ML) algorithms. This project focuses on both the development of hybrid methods coupling CFD and ML, and their implementation on relevant industrial applications.

CFD will be carried out with Fluidity, which is an open source multi-phase CFD code capable of numerically solving the Navier-Stokes and field equations. Fluidity uses a moving Finite Element/Control Volume method which enables anisotropic mesh adaptivity on unstructured meshes for time dependant problems (AMCG, 2015). To modify the mesh, Fluidity uses hr-adaptivity, which is a combination of h-adaptivity and r-adaptivity, with the former changing the connectivity of the mesh and the latter relocating its vertices. In practice, this mean that the resolution can be increased or decreased in certain areas depending on the field of interest, e.g. pressure or velocity. For example, in a multiphase flow with a moving interface, a high resolution can be obtained near the interface as it moves and a lower resolution further away from it. In the case of flow past a body, this method can maintain a high mesh resolution in turbulent areas in the wake while retaining a coarse mesh in the far field. This results in improved computational and storage savings compared to fixed mesh methods.

In recent years, ML has been applied to a range of CFD applications. A major field of interest is applying ML to turbulence closure models. Beck et al. (2019) used Deep Neural Networks (DNN) to develop subgrid-scale (SGS) models by training them on DNS data, as opposed to physics based SGS models. While it was found that purely data-based SGS models were not applicable to practical applications, they confirmed that "data-informed" closure models have potential and suggest further work in that area.
Another particularly relevant application with multiphase flows is through the use of ML classification algorithms. Guillén-Rondon et al. (2018) used support vector machines (SVM) to predict flow patterns in two-phase gas-liquid flow, using
experimental training data.

The project will be entirely computational and will aim to relate ML and CFD for multiphase flow problems. This will be achieved through the simulation of multiphase flows using Fluidity, the implementation of ML algorithms with scikit-learn, and the development of codes coupling the two together.