UC San Diego

Steady State fluid simulations are a critical piece of the mechanical engineering design loop but serve as a bottleneck due to the engineering overhead and the high computational load. In recent years, there has been an increase in the research activity in the field of neural fluid simulations, however, the current works have limited scope with respect to the fluid domains and flow regimes, restricting their generalizability and their potential industrial impact. This thesis seeks to introduce a foundational framework for generic 3D steady state neural fluid simulations, with the goal of simplifying and expediting the fluid simulation process for engineering applications. To that end, this project introduces 3 key contributions: A python package for generating and post-processing fluid simulations in Ansys in an automated end-to-end fashion, a benchmark dataset of over 3060 fluid channel geometries and roughly 3400 quality fluid simulations in line with engineering standards for accuracy, and a series of foundational experiments exploring steady state neural fluid simulations for internal channel flows - the first of its kind as far as the author is aware. Experimental results demonstrate that geometric deep learning models have the capacity to be used as a proxy for traditional fluid simulations, but also indicate that further research is required to continue to develop the datasets and architectures for this deep learning application.