An Application of Generative Adversarial Networks for Robust Inference in Computational Fluid Dynamics

In this paper we propose a robust learning pipeline for inference in computational fluid dynamics (CFD) systems in the presence of faulty sensor data. The standard methods for handling faulty sensor data involve outlier detection techniques which assume that the faulty data is generated from the tail regions of the underlying data distribution and hence can be eliminated by modeling the high probability regions of the distribution. However this assumption is not always true and subtle faults in sensors can lead to recording of faulty data which can be thought of as being generated from a subtly perturbed version of the underlying distribution. Methods based on outlier detection techniques will fail to work under these settings and hence novel approaches are required for eliminating faulty data in such systems. In this work we explore the use of a Generative Adversarial Network (GAN) for this purpose. We train the generator network of the GAN to generate "fake" sensor data that mimics the distribution of the real data, albeit, a slightly perturbed one. We use this to train a discriminator network which learns to distinguish between the "real" and "fake" data generated from the generator. This discriminator is then used to filter out faulty sensor data generated from a perturbed version of the distribution generating the real data. We also build a simple regressor that uses the trained discriminator to perform robust regression on the CFD data after eliminating faulty sensor data. We tested the robust regression pipeline with CFD data for predicting fluid flow characteristics (specifically the angle of attack (AoA)) over a 2D foil. Our discriminator trained in a GAN framework could eliminate faulty sensor data, generated using the trained generator, with ∼ 100 % efficiency. The filtered data is then used for inference of the fluid flow parameters using the regressor.