Hybrid Techniques for Data-Physics Based Machine Learning in Computational Fluid Dynamics
Loading...
Authors
Aulakh, Deepinder
Date
Type
thesis
Language
eng
Keyword
Machine Learning , CFD , RANS
Alternative Title
Abstract
The application of physics-based machine learning (PBML) in fluid dynamics is limited by slow convergence and over-fitting problems. These difficulties arise mainly from the limitations of current PBML methods in handling coupled partial differential equations, such as the Navier-Stokes equations. To overcome these issues, this thesis proposes a solution by integrating PBML with principles derived from computational fluid dynamics (CFD). The proposed approach uses numerical/discrete loss functions and integrates them with pressure-linked algorithms of CFD. This approach is successfully applied to equations such as Poisson and Navier-Stokes equations, selected for varying degrees of mathematical complexity. The thesis also explores the unification of physics and data-driven learning, enabling low-quality data to be used for accurate and accelerated training of neural networks. Two physics-data integration approaches, namely `hybrid' and `mixed' learning, are proposed in this thesis. These approaches are successfully tested on sparse and noisy (low-quality) data. The proposed methods are implemented in an integrated CFD-PBML application specifically developed for PBML, providing a comprehensive testing framework for future applications. Together, the results of the thesis highlight the potential of CFD-PBML to tackle complex learning tasks of fluid mechanics.