Data-driven subgrid scale modelling with neural networks
An exploratory study is performed to assess the proficiency of the neural networks in prediction of the non-linear mapping of the closure terms of LES and the coarse grid components in the flow, with \textit{a priori} assumptions. Two distinct frameworks of convolutional neural networks are built to...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2020
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/137038 |
| _version_ | 1826126294865149952 |
|---|---|
| author | Gangu,Vaishnavi |
| author2 | Ng Bing Feng |
| author_facet | Ng Bing Feng Gangu,Vaishnavi |
| author_sort | Gangu,Vaishnavi |
| collection | NTU |
| description | An exploratory study is performed to assess the proficiency of the neural networks in prediction of the non-linear mapping of the closure terms of LES and the coarse grid components in the flow, with \textit{a priori} assumptions. Two distinct frameworks of convolutional neural networks are built to interpret the relation between the subgrid scale stress and the filtered velocity components. The first approach being the super resolution convolutional neural networks (SRCNN), originally a design for image super resolution, is found to reconstruct the high resolution flow field with a remarkable level of accuracy. Subsequent measures involved the extraction of SGS stress from this high resolution flow field. The second framework involves a direct prediction of the SGS behaviour from the filtered velocity components, exhibiting satisfactory performance. Additional examination of the model architecture encompassed the altering of the convolution kernel width of the intermediate layer. With positive and favourable results, the proposed convolutional neural network frameworks could establish a foundation for the development of potential data-driven subgrid scale models of more complex turbulent flows. |
| first_indexed | 2024-10-01T06:50:20Z |
| format | Thesis-Master by Coursework |
| id | ntu-10356/137038 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T06:50:20Z |
| publishDate | 2020 |
| publisher | Nanyang Technological University |
| record_format | dspace |
| spelling | ntu-10356/1370382023-03-11T17:47:42Z Data-driven subgrid scale modelling with neural networks Gangu,Vaishnavi Ng Bing Feng School of Mechanical and Aerospace Engineering École Polytechnique Fédérale de Lausanne Technical University of Munich bingfeng@ntu.edu.sg Engineering::Aeronautical engineering::Aerodynamics An exploratory study is performed to assess the proficiency of the neural networks in prediction of the non-linear mapping of the closure terms of LES and the coarse grid components in the flow, with \textit{a priori} assumptions. Two distinct frameworks of convolutional neural networks are built to interpret the relation between the subgrid scale stress and the filtered velocity components. The first approach being the super resolution convolutional neural networks (SRCNN), originally a design for image super resolution, is found to reconstruct the high resolution flow field with a remarkable level of accuracy. Subsequent measures involved the extraction of SGS stress from this high resolution flow field. The second framework involves a direct prediction of the SGS behaviour from the filtered velocity components, exhibiting satisfactory performance. Additional examination of the model architecture encompassed the altering of the convolution kernel width of the intermediate layer. With positive and favourable results, the proposed convolutional neural network frameworks could establish a foundation for the development of potential data-driven subgrid scale models of more complex turbulent flows. Master of Science (Aerospace Engineering) 2020-02-13T07:02:04Z 2020-02-13T07:02:04Z 2019 Thesis-Master by Coursework https://hdl.handle.net/10356/137038 en application/pdf Nanyang Technological University |
| spellingShingle | Engineering::Aeronautical engineering::Aerodynamics Gangu,Vaishnavi Data-driven subgrid scale modelling with neural networks |
| title | Data-driven subgrid scale modelling with neural networks |
| title_full | Data-driven subgrid scale modelling with neural networks |
| title_fullStr | Data-driven subgrid scale modelling with neural networks |
| title_full_unstemmed | Data-driven subgrid scale modelling with neural networks |
| title_short | Data-driven subgrid scale modelling with neural networks |
| title_sort | data driven subgrid scale modelling with neural networks |
| topic | Engineering::Aeronautical engineering::Aerodynamics |
| url | https://hdl.handle.net/10356/137038 |
| work_keys_str_mv | AT ganguvaishnavi datadrivensubgridscalemodellingwithneuralnetworks |