Intelligent modelling in Computational Rheology: introduction of Machine Learning approaches and exploitation

点击翻译按钮获取中文摘要

In recent years, a vast array of industrial disciplines dealing with complex fluid systems have become eager for numerical methods and approaches which can provide quantitative numerical results to enhance industrial planning and to minimise the expenses resulting from trial-and-error procedures and attempts to gather empirical data. Companies in various engineering fields from oil and gas and bioengineering, up to food production, either utilise or create products possessing complex rheological characteristics such as viscoelasticity, and their behaviours deviate significantly from the simple and well-understood characteristics of Newtonian fluids such as water. Their peculiar flow behaviours arise from fundamentally different microstructures, resulting in complex mechanical properties.When performing numerical simulations to approximate real fluid flows it is important to quantify precisely the underlying physical mechanisms for predicting as accurate as possible the real-world behaviour. Achieving a full resolution of the resulting dynamics often requires very expensive computational procedures that demand time. An additional complexity is induced when the problems under investigation are in the field of Computational Rheology (CR) where non-Newtonian viscoelastic fluids are considered. In contrast to the "simple" fluids that obey Newton's law of viscosity, the models employed for investigating these complex fluids require additional equations that need to be solved and exhibit strongly non-linear responses, which enormously increase the computational demands. The majority of models employed in CR originate from kinetic theories and series of closures have been proposed for deriving closed-form equivalents for use in Computational Fluid Dynamics (CFD). Furthermore, to approximate experimental responses, most of these models introduce physical variables which are typically treated as constants and are responsible for controlling the rheological properties of the simulated fluids. All the above, apply drastic simplifications to the rich dynamics that are imposed by the complex fluid microstructure and thus, when are employed in CFD they cannot produce accurate results from a quantitative perspective. Attempts made to increase the accuracy of the closed-form models, led to a greater complexity, where new, non-linear, constitutive relationships are proposed which introduce additional parameters that need to be controlled. Results are encouraging in terms of accuracy, where for a good range of parameters the experimental measurements are more closely approximated, but still are far from real observations and responses.This PhD project is aiming to develop further novel numerical models and procedures, attempting to increase the accuracy and predictivity of CFD calculations when investigating flows of non-Newtonian fluid systems and provide new results and insights of behaviours met in complex fluid flows. Modifications will be applied to the promising closed form of the Adaptive Length Scale model and Machine Learning approaches will be incorporated in CR which are the future in CFD applications and numerical modelling in fluid mechanics, aiming to shape new routes in fluid flow simulations.