RANN: A Neural RANS Closure Model for Physics-Informed Machine Learning on General Geometries

Aerodynamic design involves the use of various techniques to optimize the geometry of a design component, to achieve desired characteristics such as increased lift or decreased drag. This usually requires iterative design loops that are slow and computationally expensive, and may rely on simplified turbulence models such as Reynolds-Averaged Navier Stokes (RANS).

We propose the Reynolds-Averaging Neural Network (RANN), a neural network based RANS closure model to speed up these design loops. RANN is able to produce a fast inference model for RANS solutions which produces accurate results on unseen geometries nearly instantly. We leverage time-tested fluid dynamics modeling ideas in a physics-informed machine learning regime, combined with novel deep learning architectures to enable generalization to geometries outside the training set. This allows practitioners to quickly and exhaustively explore the design space of an aerodynamic optimization problem. We present an implementation of the neural inference model, a demonstration of the results, and an empirical comparison to state of the art methods.