Towards Building Robust Neural Network Models for Fluid Simulations

The rise of Machine Learning (ML) for modeling complex problems in fluid physics has brought with it challenges in developing tools that are robust and explainable. Scientific data being high-dimensional, multimodal, and complex makes it difficult to choose the appropriate hyper-parameters for such networks. In this work, we explore the possibility to fully automate the network design process for a data-driven fluid physics emulator - in this case modeling a key turbulence prognostic critical for closure in a Large Eddy Simulation (LES) compressible flow code; including neural architecture search as well as optimizing the hyper-parameters using Bayesian principles. Further we investigate the network learnings by analyzing the gradient flows and conduct variance-based sensitivity analysis to understand the trained network predictions thereby improving explainability of these so called black-box models