This title appears in the Scientific Report :
2020
Please use the identifier:
http://hdl.handle.net/2128/24572 in citations.
Sub-Grid Scale Modelling at Scale with Deep Learning and up to 60 Billion Degrees of Freedom
Sub-Grid Scale Modelling at Scale with Deep Learning and up to 60 Billion Degrees of Freedom
This work presents fully resolved direct numerical simulations (DNSs) of a turbulent reactive planar temporally non-premixed jet configuration with up to 60 billion degrees of freedom. As scalar mixing is of utmost importance for this kind of configuration, a novel deep learning (DL) approach in the...
Saved in:
Personal Name(s): | Bode, Mathis |
---|---|
Denker, Dominik / Jitsev, Jenia (Corresponding author) / Pitsch, Heinz | |
Contributing Institute: |
John von Neumann - Institut für Computing; NIC Jülich Supercomputing Center; JSC |
Published in: |
NIC Symposium 2020 |
Imprint: |
Jülich
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
2020
|
Physical Description: |
379 - 388 |
Conference: | NIC Symposium 2020, Jülich (Germany), 2020-02-27 - 2020-02-28 |
Document Type: |
Contribution to a book Contribution to a conference proceedings |
Research Program: |
Data-Intensive Science and Federated Computing |
Series Title: |
Publication Series of the John von Neumann Institute for Computing (NIC) NIC Series
50 |
Link: |
OpenAccess OpenAccess |
Publikationsportal JuSER |
This work presents fully resolved direct numerical simulations (DNSs) of a turbulent reactive planar temporally non-premixed jet configuration with up to 60 billion degrees of freedom. As scalar mixing is of utmost importance for this kind of configuration, a novel deep learning (DL) approach in the context of large-eddy simulation is presented which results in predictive mixing statistics on underresolved grids. The usability of the mixing model is approved by applying it to the DNS data. Furthermore, node performance measurements for the training of the DL networks are shown for different computing clusters. |