Gene Expression Programming (GEP), a branch of machine learning, is based on the idea to iteratively improve a population of candidate solutions using an evolutionary process built on the survival-of-the-fittest conce...
详细信息
Gene Expression Programming (GEP), a branch of machine learning, is based on the idea to iteratively improve a population of candidate solutions using an evolutionary process built on the survival-of-the-fittest concept. The GEP approach was initially applied with encouraging results to the modeling of the unclosed tensors in the context of RANS (Reynolds Averaged Navier-Stokes) turbulence modeling. In a subsequent study it was demonstrated that the GEP concept can also be successfully used for modeling the unknown Sub-Grid Stress (SGS) tensor in the context of Large Eddy Simulations (LES). This was done in an a-priori analysis, where an existing Direct Numerical Simulation (DNS) database was explicitly filtered to evaluate the unknown stresses and to assess the performance of model candidates suggested by GEP. This paper presents the next logical step, i.e. the application of GEP to a-posteriori LES model development. Because a-posteriori analysis, using in-the-loop optimization, is considered the ultimate way to test SGS models, this can be considered an important milestone for the application of machine learning to LES based turbulence modeling. GEP is here used to train LES models for simulating a Taylor Green Vortex (TGV) and results are compared with existing standard models. It is shown that GEP finds a model that outperforms known models from literature as well as the no-model LES. Although the performance of this best model is maintained for resolutions and Reynolds numbers different from the training data, this is not automatically guaranteed for all other models suggested by the algorithm. (C) 2020 Elsevier Inc. All rights reserved.
暂无评论