Enhancing machine learning turbulence models via prediction of tensor basis normalized coefficients

• Main Authors: Ji, Ziqi, Duan, Penghao, Du, Gang
• Format: Journal Article
• Language: English
• Published: 16.10.2024
• Subjects: Physics - Fluid Dynamics
• DOI: 10.48550/arxiv.2410.12255

With the rapid advancement of machine learning techniques, their application in developing turbulence models has become increasingly prevalent. Numerous machine learning turbulence models have recently been founded on Pope's general eddy-effective hypothesis. However, most of these models directly predict the non-dimensional Reynolds deviatoric tensor, treating tensor basis coefficients as intermediate variables. This approach presents three primary drawbacks. Firstly, it conflates the training error of the machine learning regression algorithm with the modeling error inherent in Pope's general eddy-effective hypothesis. Secondly, by treating tensor basis coefficients as intermediary values, researchers often overlook their significance and physical interpretation, particularly in black-box models such as neural networks. Thirdly, this method introduces numerical interdependence among tensor basis coefficients, violating the orthogonality of the tensor basis. Yin et al. (2022) proposed a neural network-based turbulence model that utilizes tensor basis coefficients as prediction labels. However, their framework employs a regularization method with numerous hyperparameters to achieve convergence and smooth results by leveraging information from the linear-eddy viscosity hypothesis. In this study, we introduce a novel approach to enhance the accuracy and realizability of machine learning turbulence models by predicting tensor basis normalized coefficients. This method's fundamental principle is the normalization and bounding of prediction labels for machine learning algorithms. Utilizing periodic hills direct numerical simulation (DNS) data for training and testing, we demonstrate that our innovative method improves accuracy and realizability in neural network-based and symbolic regression-based turbulence models.