Data-Driven Modelling of Multiphase Flow Systems

Dynamical systems specifically in the field of fluid mechanics are composed of underlying complicated governing phenomena originated from nonlinearities and instabilities. Encountered with the challenge of analyzing vast amount of data, the concept of reduced order modelling (ROM) was emerged to map the high resolution spatio-temporal data onto a low-dimensional space using the most prominent embedded features. This dissertation considers two ROM techniques of proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) applied to liquid injection systems. These approaches have been widely used to tackle the challenges of analyzing spatio-temporal coherence of dynamical systems. Despite the numerous works implementing POD and DMD, there has been a lack of physical meaning for the modes generated by them. An interpretation of POD and DMD modes is provided in this thesis by the recognition of dominating features. The main focus will be primitively on benchmark problems to validate the efficacy of the methods and consequently to the liquid jets exposed to air crossflows in a hierarchical scheme. A grasp of the prominent spatial structures and their corresponding leading dynamic frequencies will be provided through the analysis of POD and DMD frequency spectra. Effects of several different factors such as the gaseous Weber number, liquid-gas momentum flux ratio and the injector aspect ratio are investigated in this study. Finally, the power of ROM techniques to create features for machine-learnt classifiers that are sufficient for categorization of sundry types of flow regimes is investigated in a supervised manner. These classifiers are opted from a range of classical machine learning algorithms like support vector machines (SVM) and random forest (RF) that have been extensively employed for classification tasks in the recent years. The best combination of reduced order models with the machine learning algorithms are presented