Publications of the research results / Documentation on the approach of the research group / Tutorials

Zahn, R., Linke, T., Breitsamter C.: Neural Network Modeling of Transonic Buffet on the NASA Common Research Model. NNFM, Vol. 151, 2021, pp. 698-709.

Rozov, V., Breitsamter, C.: Data-driven prediction of unsteady pressure distributions based on deep learning. Journal of Fluids and Structures, Vol. 104, 2021. https://doi.org/10.1016/j.jfluidstructs.2021.103316

Andre Weiner and Richard Semaan: flowTorch - a Python library for analysis and reduced-order modeling of fluid flows, Journal of Open Source Software, 6(68), 3860, 2021, https://doi.org/10.21105/joss.03860

Zahn, R., Breitsamter, C.: Airfoil buffet aerodynamics at plunge and pitch excitation based on long short-term memory neural network prediction.  CEAS Aeronaut J 13, 45–55 (2022).  https://doi.org/10.1007/s13272-021-00550-6

Zahn, R., Breitsamter C.: High-speed buffet aerodynamics modeling based on a long short-term memory neural network. DLRK2020, Paper 0027, 2020.

Zahn, R., Breitsamter, C.: Prediction of Transonic Wing Buffet Pressure Based on Deep Learning. DLRK2021, Paper 0033, 2021.

Daniel Fernex, Andre Weiner, Bernd Noack and Richard Semaan: Sparse Spatial Sampling: A mesh sampling algorithm for efficient processing of big simulation data, AIAA 2021-1484, AIAA Scitech 2021 Forum, January 2021, https://doi.org/10.2514/6.2021-1484

Andre Weiner and Richard Semaan: Simulation and modal analysis of transonic shock buffets on a NACA-0012 airfoil, AIAA 2022-2591, AIAA SCITECH 2022 Forum, January 2022, https://doi.org/10.2514/6.2022-2591

The flowTorch library [1, 2] enables researchers to access, analyze, and model fluid flow data from experiments or numerical simulations. Instead of a black-box end-to-end solution, flowTorch provides modular components allowing to assemble transparent, automated, and reproducible processing workflows with ease. Fluid flow data coming from experiments (PIV, iPSP, Schlieren) or simulations (TAU, Flexi, OpenFOAM) may be accessed via a common interface in a few lines of Python code. Internally, the data are organized as PyTorch tensors. Relying on PyTorch tensors as primary data structure enables fast array operations, parallel processing on CPU and GPU, and exploration of novel deep learning-based analysis and modeling approaches. The flowTorch documentation includes a rich collection of Jupyter notebooks [3] demonstrating how to apply the library components in a variety of different use cases, e.g., finding coherent flow structures with modal analysis or creating reduced-order models.

The image shows the dominant pattern (mode) in the unsteady surface pressure distribution on the XRF1 model. Both proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) are available in flowTorch for the analysis of unsteady flows.

The DMD decomposes the flow into several characteristic spatial flow structures, so-called modes. Each mode is associated with a unique frequency. By superposition of a few dominant modes close to the buffet frequency, the main flow features can be recovered, as seen in the animation.

[1] https://doi.org/10.21105/joss.03860

[2] https://github.com/FlowModelingControl/flowtorch

[3] https://flowmodelingcontrol.github.io/flowtorch-docs/1.0/index.html