Training of neural networks using the example of a heat exchanger based on existing CFD data
The aim of this example is to use a trained ANN to determine the geometry variant for a new heat exchanger in a matter of seconds in the future, which provides the best possible cooling capacity, without having to perform complex CFD simulations again. An ANN was trained based on a large number of existing simulation results for different geometry variants.
Subsequent presents the ANN training process using the existing CFD heat sink simulations database. Tensorflow was used as the platform for the ANN training and Anaconda as the software environment.
Approximately 800 data sets were used for the present ANN. Each data set contains geometric and physical parameters of the respective heat sink as well as the result of the respective CFD simulation (cooling capacity). 70% of the existing datasets were used for training and for testing and validating the trained ANN respectively 15%.
After successful ANN training, the cooling capacity for varying input data of a heat sink can be predicted with an accuracy of +/-10%.
"Computers are just shocking consistent."