Options
Neural networks based processing of CFD data for supporting the early design of ship propellers
Citation Link: https://doi.org/10.15480/882.9314
Publikationstyp
Conference Paper
Publikationsdatum
2024-04-04
Sprache
English
Start Page
507
End Page
514
Citation
8th International Symposium on Marine Propulsors (smp 2024)
Contribution to Conference
Publisher
Norwegian University of Science and Technology, Department of Marine Technology
ISSN
2414-6129
ISBN
978-82-691120-5-4
Peer Reviewed
true
While powerful and reliable CFD methods are available today, designing propellers using them is unfortunately timeconsuming and cost-intensive. On the contrary, existing CFD results for the hydrodynamic behaviour of propellers can be stored and processed with neural networks. This paper deals with the acceleration of the early propeller design stage by using this data for the training of neural networks.
For the generation of the dataset for the network training different radial propeller parameters including skew, chord length, camber and pitch for three to six bladed propellers were developed. The outputs of the automatic CFD calculations are the pressure distribution over the blade surface and the thrust and moment coefficients for different advance ratios.
Then all parameters are processed in a two-step neural network; the first neural network processes the operating point data and the second network the geometry data and estimated cavitation area. Based on the requirements such as installed power or the required propeller thrust the optimal geometry regarding a maximized efficiency is determined.
The grid sensitivity study in combination with processing different amounts of training data show that with sufficient data quality, a higher amount of data reduces the averaged error values in both networks until a certain point.
For the generation of the dataset for the network training different radial propeller parameters including skew, chord length, camber and pitch for three to six bladed propellers were developed. The outputs of the automatic CFD calculations are the pressure distribution over the blade surface and the thrust and moment coefficients for different advance ratios.
Then all parameters are processed in a two-step neural network; the first neural network processes the operating point data and the second network the geometry data and estimated cavitation area. Based on the requirements such as installed power or the required propeller thrust the optimal geometry regarding a maximized efficiency is determined.
The grid sensitivity study in combination with processing different amounts of training data show that with sufficient data quality, a higher amount of data reduces the averaged error values in both networks until a certain point.
Schlagworte
propeller design
neural networks
computational fluid dynamics
DDC Class
620: Engineering
Publication version
publishedVersion
Loading...
Name
Strecker-NeuralNetworksBasedProcessingOfCfdDataForSupportingTheEarlyDesig-1126-1-final.pdf
Type
main article
Size
1.83 MB
Format
Adobe PDF