Wave Inversion with Machine and Deep Learning

Author: Maximilian Schmitz @ Georgia Institute of Technology

Email: mschmitz7@gatech.edu

Introduction

This is the official repository of the publication

Schmitz, Maximilian and Kim, Jin-Yeon and Jacobs, Laurence J., Machine and Deep Learning for Coating Thickness Prediction Using Lamb Waves. Available at ScienceDirect: here or http://dx.doi.org/10.2139/ssrn.4124083

With this repository, you can

• convert phase speed - frequency representations into the frequency - wave number domain
• automatically simulate CAE models for finite element analysis in Abaqus
• invert to the thickness of the coating using machine learning techniques
• invert to the uniformness of the coating using deep learning

Repo Walk through

The project is subdivided into 3 separate main parts. Each part is contained in a separate folder with respective naming which will be walked through in the following. There is a visualization of the repo structure below too.

• analysis_sklearn contains the machine learning inversion procedure based on scikit-learn
• analysis_disp_curves incorporates the conversion of dispersion curves extracted from a dispersion curve calculator (like DLR dispersion calculator) into the right format and domain, e.g. from a phase speed - frequency representation into a frequency - wave number domain.
• create_model includes all scripts and files to (1) create the CAE model with a Python script in Abaqus (industry standard for finite element analysis (FEM)) and (2) to run the entire pipeline for creating several simulations on a Linux system (like the Georgia Tech PACE cluster) in parallel. More detailed information about how to use those scripts properly and how to install the respective environment can be found in the README.md in the respective folder.
• dl_code contains the code and data for training and evaluating a deep learning based inversion model. The folder contains several notebooks which are explained in the README.md.
• extract_disp_history is merged into the create_model folder and is not needed anymore
• helper_scripts include smaller scripts for minor tasks used during develpment
• unit_tests include Python unit tests which were run during development and can still be used for testing any adjustments made. Most unit tests are called form the deep learning code.

Repo Structure

ml_dl_wave_inversion
│   README.md  
│
└───analysis_sklearn
│   │   analysis_sklearn.ipynb
│   │   dispersion_feature_loader.py
│   └───data
|       |   ...
│   └───figures
│
└───analytical_disp_curves
│   |   analytical_curves.py
│   └───Cr_dispersion_curves
|      |   ...
│   └───Zy4Cr_dispersion_curves
|      |   ...
│   └───Zy4_dispersion_curves_DC
|      |   ...
│
└───create_model
│   |   README.md
|   |   __init__.py
|   |   at_cluster_conda.pbs
|   |   at_cluster_conda_postpro.pbs
|   |   auto_feat_ext.pbs
|   |   auto_feature_extraction.py
|   |   auto_image_extraction.py
|   |   auto_img_ext.pbs
|   |   create_model_script_v15.py
|   |   create_model_script_v15_no_gap.py
|   |   delete_unnecessary_files_from_cluster.py
|   |   extract_disp_history_max_v5.py
|   |   postprocess_2dfft_max_v15.py
|   |   qdel_automation.py
|   |   run_automated_simulations_cluster.py
|   |   run_automated_simulations_cluster_postp.py
|   |   run_parallel_on_cluster.pbs
|   |   run_parallel_simulations.py
|   |   run_simulation.py
|   |   utils.py
|   |   visualize_dataset.py
|   |   wave_env_cluster.yml
│   └───figures_param_space
|
└───dl_code
│   |   README.md
|   |   __init__.py
|   |   confusion_matrix.py
|   |   data_transforms.py
|   |   dl_utils.py
|   |   image_loader.py
|   |   maxWaveNet_local.ipynb
|   |   my_resnet.py
|   |   my_resnet_old_december.py
|   |   optimizer.py
|   |   runner.py
|   |   simple_net.py
|   |   simple_net_big.py
|   |   simple_net_final.py
|   |   stats_helper.py
|   |   test_cuda_on_cluster.ipynb
|   |   testing.py
|   |   train_network.pbs
|   |   train_network_cluster.py
|   |   validate_network.ipynb
|   |   visualize_training_history.py
|   |   wave_cnn_env_linux.yml
|   |   wave_cnn_env_linux_2.yml
│   └───data
|      └───test
|           |   ...
|      └───train
|           |   ...
│   └───figures
│   └───model_checkpoints
│   └───trained_models
│
└───helper_scripts
│   |   automate_calling_from_cmd.py
|   |   build_slack_notifications.py
│   |   calculate_group_velocity.py
|   |   calculate_material_constants.py
│   |   delete_only_blacklist_files.py
|   |   ft_triangular_func.png
│   |   triangle_fourier_transform.py
|   
└───resources
| 
└───unit_tests
│   |   __init__.py
|   |   model_test_utils.py
│   |   test_data_transforms.py
│   |   test_dl_utils.py
|   |   test_image_loader.py
│   |   test_my_resnet.py
│   |   test_simple_net.py
|   |   test_simple_net_final.py
│   |   test_stats_helper.py
|   |   utils.py

Train/Test Data & Pretrained Weights

As by definition, the machine and deep learning part is based on simulated data. I have not uploaded the data used for this research (train and test images, pretrained model parameters) to providers like SourceForge, ... yet. If you are interested, feel free to send me an email and I am happy to provide you with the respective data: mschmitz7@gatech.edu

Citation

If you find this useful, please cite our paper.

@article{ml_wave_inversion,
 title = {Machine and deep learning for coating thickness prediction using Lamb waves},
 author = {Maximilian Schmitz and Jin-Yeon Kim and Laurence J. Jacobs},
 journal = {Wave Motion},
 volume = {120},
 pages = {103137},
 year = {2023},
 issn = {0165-2125},
 doi = {https://doi.org/10.1016/j.wavemoti.2023.103137},
 url = {https://www.sciencedirect.com/science/article/pii/S0165212523000239},
 keywords = {Machine learning, Deep learning, Inversion, Wave propagation, Lamb wave, Fourier transform},
 year={2022}
}