A Comparison of Reduced-Order Models for Wing Buffet Predictions

Abstract

The primary aim of this study is to train and evaluate a set of Reduced-Order Models (ROMs) for predicting upper-wing pressure distributions on a civil aircraft configuration. Using wind tunnel data recorded for the Airbus XRF-1 research configuration in the European Transonic Windtun- nel (ETW), the ROMs integrate dimensionality reduction and time-evolution components. For dimensionality reduction, both Singular Value Decomposition (SVD) and a convolutional Vari- ational Autoencoder (CNN-VAE) neural network are employed to reduce/encode Instationary Pressure Sensitive Paint (IPSP) data from a transonic buffet flow condition. Fully-Connected (FC) and Long Short-Term Memory (LSTM) neural networks are then applied to predict the evolution of the latent representation, which is subsequently reconstructed/decoded back to its original high-dimensional state.

SVD and CNN-VAE are trained on data from five flow conditions and tested on two unseen conditions. When comparing the power spectra of the reconstructed and experimental data, SVD demonstrates marginally better performance. Subsequently, FC and LSTM models are applied for the forward evolution of the reduced/latent representation for one flow condition, resulting in four evaluated ROMs. Among them, the CNN-VAE-LSTM model excels by accurately capturing buffet dynamics, encompassing both transient features and steady-state oscillations. SVD-based models tend to struggle with transient buffet behavior, as they predominantly learn steady-state dynamics. Additionally, the CNN-VAE-LSTM model was employed for an end-to-end training approach. The results suggest that it faces difficulties maintaining dynamics in predictions, demanding further investigation and optimization.

BibTex Citation

@phdthesis{schreiber2023,
  author       = {Schreiber, Anton},
  title        = {{A Comparison of Reduced-Order Models for Wing 
                   Buffet Predictions}},
  school       = {Technische Universität Braunschweig},
  year         = 2023,
  month        = dec,
  doi          = {10.5281/zenodo.10255520},
  url          = {https://doi.org/10.5281/zenodo.10255520}
}

Dependencies

Python environment

To set up a suitable virtual environment, execute the following commands:

sudo apt install python3-venv
python3 -m venv ipsp
# activate the environment
source ipsp/bin/activate
# install dependencies
pip install -r requirements.txt
# leave the environment
deactivate

Data

The datasets are expected to be located in a folder named data, which is ignored by the version control system.

Getting started

A typical workflow might look as follows:

• Load your data with .pt format into the data folder with:
  1. a $c_p$-snapshot file containing different flow conditions, that should be stored in a dictionary. The keys represent the flow condition name and the corresponding values represent the pressure data in form of PyTorch Tensors with (height, width, number of snapshots)
  2. a coordinate or mesh file, that could also be stored in similar dictionary format. The keys also represent the flow condition, while the values are tuples containing coordinate meshes for the height and the width dimension

(Note for all following steps that you probably need to change path names in the scripts and adjust parameters in the app/utils/config.py file according to your specific data)

• Interpolate your data and limit the number of timesteps with the interpolate_coords and make_data_subset functions in app/preprocessing.py. This is optional but can significantly reduce required computational ressources.
• In app/preprocessing.py, use svd_preprocesing and autoencoder_preprocessing to create datasets for the training of SVD and CNN-VAE.
• In the next step, you can train the dimensionality reduction techniques:
  • SVD: Simply run the SVD_train.ipynb file in the notebooks directory.
  • CNN-VAE: The training of a CNN-VAE neural network is more complex. Head to autoencoder directory in app and open train_VAE.py. This file is used to start a parameter study for different bottleneck layer sizes of the CNN-VAE models. All training parameters are specified in the config.py file, adjust them to your needs (this also applies for all other neueral network trainings)
• To monitor the training, use the corresponding <>_optim.ipynb notebooks in the notebooks directory. To apply both techniques to test data, use the corresponding <>_test.ipynb notebooks. Adjust parameters and repeat the training (optional).
• Similarly, FC and LSTM neural networks can be trained to auto-regressively predict the time-series in the reduced/ latent space. In app, there are FC and LSTM directories where you can find the model classes and the training scripts. There, you can set the dimensionality reduction technique to pair the model with and start a training pipelines. In the frame of the course project, they were trained on a single flow condition using the single_flow_cond_preprocessing function in the preprocessing.py. You can also copy and adjust the scripts to train with multiple flow conditions, the pre-processing scripts are already available. The only difference is that you should consider a 3rd dataset when dealing with multiple flow conditions to end up with a training, a validation and a test dataset, which is not sensible for a single flow condition.
• In the last step, use corresponding <>_optim.ipynb and <>_test.ipynb notebooks to monitor the training and test the models on unseen data.

Extensibility

This repository can be extended by adding new dimensionality reduction and time-evolution techniques in the same modular structure.