Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Using Neural Networks to Compute the Divergence of the Reynolds Stress Tensor with Fundamental Mean Flow Properties
Submission Author:
Thales Arantes de Castelo Branco e Souza , SP
Co-Authors:
Thales Arantes de Castelo Branco e Souza, William Wolf, Luiz Augusto Camargo Aranha Schiavo
Presenter: Thales Arantes de Castelo Branco e Souza
doi://10.26678/ABCM.COBEM2023.COB2023-0079
Abstract
Numerical simulations of complex flows commonly rely on the solution of the Reynolds-Averaged Navier-Stokes (RANS) equations due to the high computational cost of high-fidelity simulations. However, RANS simulations model the entire range of turbulent scales, requiring closure models for the Reynolds stresses. This paper introduces an alternative approach that avoids the use of turbulence models for correcting the Reynolds stress field. Instead, it employs machine learning techniques based on neural networks (NNs) to directly calculate the divergence of the Reynolds stress tensor using mean flow properties (pressure, velocity, and their gradients) which are readily available from a RANS calculation. The proposed NN model successfully reproduces the divergence of the Reynolds stress tensor in a turbulent flow within a convergent-divergent channel, where separation and reattachment occur downstream of a smooth bump. The strong physical correlation between the mean properties and the divergence of the Reynolds stress tensor enables the machine learning solution to be data-efficient and computationally inexpensive. Accurate reconstructions of the divergence of the Reynolds stress tensor are achieved with only 20% of the training data. Training strategies are employed to ensure rotational invariance of the solutions. Finally, the model's interpolation capacity is tested using different Reynolds numbers for training and testing, demonstrating its ability to learn the physics of turbulent flows from fundamental properties from a RANS formulation.
Keywords
Artificial Neural Networks(ANNs), Turbulence, turbulence modeling
