h1

TY  - THES
AU  - Vanvinckenroye, Joris Vincent
TI  - A Posteriori Untersuchung von CNNs zur Modellierung von Wasserstoffverbrennung
PB  - RWTH Aachen University
VL  - Masterarbeit
CY  - Aachen
M1  - RWTH-2024-08139
SP  - 1 Online-Ressource : Illustrationen
PY  - 2024
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University
N1  - Masterarbeit, RWTH Aachen University, 2024
AB  - Due to the very low pollutant emissions of hydrogen combustion, the fuel plays an important role in combating climate change and limiting global warming to 1.5C. Accurate modeling methods for hydrogen combustion are crucial to reduce the huge computational demand of direct numerical simulations and bring the technology to market maturity. In this thesis, the application of Convolutional Neural Networks, specifically U-Net and UNet++, to model lean hydrogen flames in laminar flows is investigated. We couple these networks with the physical solver CIAO using two approaches: MLLib and AIxeleratorService, and evaluate and compare their performance both in terms of accuracy and scalability. Distributed network inference using the AIxeleratorService is implemented to speed up the coupled simulation. However, we observe unphysical behavior in the coupled simulations. To combat this, we optimize the models with additional training data, hyperparameter tuning, and post-training pruning. Despite reducing the a-priori prediction errors, these methods did not consistently improve a-posteriori simulation results. Our findings emphasize the need for a-posteriori evaluation for neural networks in hydrogen combustion simulations.
LB  - PUB:(DE-HGF)19
DO  - DOI:10.18154/RWTH-2024-08139
UR  - https://publications.rwth-aachen.de/record/992307
ER  -