Fast methods for modelling fluid flow and characterising petroleum reservoirs

Abstract

This thesis tackles three kinds of computationally efficient methods widely applicable in the fields of engineering, simulation and numerical modelling. First, the Non-Intrusive Reduced Order Modelling (NIROM) is discussed, reframed, generalised and tested. While NIROM is a general methodology, the main focus of this work is to evaluate its potential in the field of reservoir modelling. For this purpose a new method for constructing parameterised NIROMs is proposed and the POD-RBF approach is investigated on a number of representative test cases. A detailed analysis concludes with NIROM not being a viable practical solution at this stage; the underlying issues, their causes and future development the method are discussed in detail.

Second, a method for classifying well log data is given. The method is an alternative to typical machine learning (ML) approaches, which up to date have been the only tools utilised for the purpose. Our approach is motivated by (and mitigates a number of) issues with applying ML in practical applications, in particular the lack of explainability. Instead of being a complex surrogate with a large number of degrees of freedom (cf ML), our model consists of the automatically re-scaled training set and a single additional number extracted during the training procedure. The technique proposed is characterised by a case-independent design, very high computational efficiency and relies on an intuitively meaningful operating principle; it also provides additional functionality in comparison with alternatives. It is demonstrated that (out of the box) the method outperforms the vast majority of alternatives on a realistic data set in terms of efficiency and accuracy, even when implemented in serial in an interpreted programming language.

Finally, the last part of the thesis addresses the issue of efficient semi-analytical modelling of solid boundaries in Smoothed Particle Hydrodynamics (SPH) simulations. More precisely, this work focuses on the purely technical aspect of efficient evaluation of correction factors governing the boundary effects; the framework utilising their values is already well established. Mathematically, the problem is described as efficiently integrating a spherically symmetric function over its compact spherical support truncated by a surface (or a collection of surfaces) representing a solid boundary (wall). Three types of boundary geometries are considered, namely piecewise-planar, spherical and super-ellipsoid/super-toroid surfaces, with the latter two categories addressed for the first time in the literature. All methods provided are characterised by an arbitrary degree of accuracy and simplicity of implementation, especially in comparison with all to up to date alternatives. A number of representative test cases is studied.