A Deep Dive into RANS Turbulence Modelling: The Role of Deep Learning in Predicting Reynolds Stress Tensor Anisotropy
Nikita Kozak, Stanford University
As a crucial foundation for numerous computational fluid dynamics (CFD) simulations, Reynolds Averaged Navier-Stokes (RANS) turbulence modelling facilitates computationally feasible solutions of complex problems that surpass the capacity of higher-fidelity approaches. Nevertheless, these RANS models are laden with substantial epistemic uncertainty, an offshoot of their overarching simplifications, informed by the universal traits observed in turbulence theory. Efforts to alleviate this uncertainty and augment RANS turbulence models have recently embraced data-driven strategies but have been confined to narrowly defined problem spaces. This present work transcends the limitations of previous studies by emphasizing the broad applicability of such data-driven RANS models, specifically for the prediction of Reynolds Stress Tensor (RST) anisotropy. We conducted an in-depth exploration of diverse deep learning topics, ranging from the implementation of domain bias and weighting frameworks to alterations in neural network (NN) design, as well as the use of coupled NNs, graph NNs, and autoencoders. This intensive analysis is conducted across a wide range of problem space complexities, providing valuable insights into the connection between a deep learning model's design and its generalizability. Our data-driven model outperforms the current state-of-the-art in predicting RST anisotropy in multifaceted problem spaces.
Abstract Author(s): Nikita Kozak, Zoe Barbeau, Gianluca Iaccarino