A New Approach for Geometric Representations in Convolutional Neural Networks for Fluid Dynamics Problems

Recent advancements in artificial intelligence have generated a lot of interest in the area of machine learning (ML) for fluid dynamics. This is important, especially for time-sensitive industrial applications, since it can provide a more efficient way of running CFD simulations, for instance, for design optimization that is inherently computationally demanding.

One class of ML algorithms commonly applied to fluid dynamics problems is the convolutional neural network (CNN). Although several different architectures of this type of neural network have been applied to fluid dynamics problems, they all rely on the basic convolutional operation. Reliance on the convolutional operation assumes that the flow domain has to be discretized using a Cartesian grid and the most common geometric representation uses the signed distance field. While this representation might be sufficient for simple problems; in situations where the geometry changes, for instance in design optimization studies, it is important to have geometric representations that can better capture the subtle differences in design options.

In this work we propose a new approach, taking into account geometric features, for representing the geometry used in CNNs applied for flow field predictions in fluid dynamics problems.