Application of AI and Machine Learning techniques for computational fluid dynamics in the Databricks cloud environment

The application of Machine Learning (ML) and Deep Learning (DL) techniques within Computational Fluid Dynamics (CFD) is increasingly moving from a frontier research topic to a concrete and practical approach. Traditional numerical methods in CFD are often hindered by high computational costs and long simulation times. ML and DL methods offer viable alternatives that not only serve as substitutes but can also enhance classical CFD approaches, addressing inherent challenges such as excessive computational demands and extended processing durations. In industrial settings, these advancements are critical for improving the efficiency and sustainability of aerodynamic, thermal, and fluid-structure interaction studies. This thesis focuses on the development of a cloud-based ML framework for CFD, designed and implemented from the ground up within the Databricks environment in collaboration with Stellantis. The approach follows a surrogate modeling paradigm, leveraging precomputed CFD data to train predictive models. Among the available methodologies, Graph Neural Networks (GNNs) were selected due to their ability to process unstructured mesh data naturally and the availability of the large-scale MegaFlow2D dataset. This dataset, comprising external flow simulations around simple geometries, provides a structured testbed for evaluating ML-based CFD surrogates. The study begins with a two-dimensional (2D) implementation, focusing on predicting flow fields around basic shapes as an initial step toward more complex scenarios. This serves as a controlled environment to refine the methodology before extending the framework to three-dimensional (3D) problems. The ultimate objective is to develop a scalable tool capable of handling more intricate geometries, aligning with real-world engineering applications where rapid and accurate fluid flow predictions are crucial.