Assessment of AI Tools for CFD Acceleration and Generalization

Abstract

Computational Fluid Dynamics (CFD) simulations are widely used in engineering applications to study fluid flow and heat transfer problems, but high fidelity simulations usually require large computational time and resources. In recent years, artificial intelligence (AI) and machine learning (ML) based surrogate models have shown strong potential for predicting CFD results using data driven approaches. These methods are becoming increasingly important in industrial applications where faster design evaluation and reduced simulation cost are desired. This thesis investigates different AI based tools and frameworks for CFD applications, including the commercial framework Altair PhysicsAI and the open source framework NVIDIA PhysicsNeMo. Different AI architectures such as Graph Neural Networks (GNNs), MeshGraphNet, Physics Informed Neural Networks (PINNs), and Fourier Neural Operators (FNOs) are studied using CFD datasets generated from ANSYS Fluent and STAR-CCM+. Two study cases are considered in the present work, namely a mixing elbow flow problem and a CPU cooling problem involving conjugate heat transfer. The main objective of the work is to evaluate the prediction accuracy, generalization capability, computational efficiency, and practical applicability of AI based surrogate models for engineering CFD problems.