Development and Implementation of Algorithms for Large-Scale CFD and Data Analytics

点击翻译按钮获取中文摘要

The term "Data Analytics" broadly refers to the multidisciplinary science of extracting meaningful conclusions from raw data. Such data-driven analyses have become an integral component in many application areas such as finance, e-commerce, weather predictions, healthcare, fraud detection, social media platforms, etc. Computational Fluid Dynamics (CFD) is one of the tools that engineers use to generate large datasets of information from which they attempt to predict the behaviour of complex fluid flows. In the last decade, CFD researchers have begun to exploit the power of data analytics in several application areas, such as turbulence modelling, uncertainty analysis and shape optimization. The need for user-friendly, reliable and computationally efficient methodologies and algorithms for seamless integration of large-scale CFD and Data Analytics has increased as industries continue to rely more heavily on numerical simulation as a tool to improve their product design. The overall aim of the proposed research is to combine the power of artificial intelligence (AI) and physics-based digital simulations to develop an innovative and robust unified framework for applying Data Analytics methods in CFD, particularly to fluid flows of industrial interest. To gain a thorough understanding and formulate an end-to-end methodology, high-fidelity CFD simulations will be conducted for several benchmark flow problems to produce typical large datasets of flow parameters. For this purpose, we will extend our research on simulating fluid flows in complex geometries using the Cartesian cut-stencil finite difference formulation of the 3D Navier-Stokes equations. Keeping in mind the desired integration of CFD and AI, our in-house CFD code will generate data in specialized formats that conform to the requirements of the AI algorithms. These datasets will be used to train and test various neural networks to access the strengths and weaknesses of different architectures. In parallel, various statistical and machine learning methods will be investigated to determine their applicability in fluid mechanics predictions and other areas of engineering science.