Adversarial Disentanglement by Backpropagation with Physics-Informed Variational Autoencoder

This repository contains the code accompanying the paper: "Adversarial Disentanglement by Backpropagation with Physics-Informed Variational Autoencoder". The pre-print version can be found here: https://arxiv.org/abs/2506.13658.

Project Structure

• cases/: Contains subdirectories for each case study (e.g., bridge, damped_oscillator, simple_beam). Each subdirectory includes the necessary models, data files, and an __init__.py file that defines the case study and its presets.
• models/: Neural network modules and model components used by the case studies.
• utils/: Utility functions, including parameter defaults, data handling, loss functions, metrics, and visualization tools. The utils/__init__.py file contains global default parameters for all case studies.
• figures/, output/: Output directories for generated figures and results.
• dpivae.py: Core functions for setting up, training, and evaluating models. This module is called by the main scripts.
• requirements.txt: List of required Python packages.

Requirements

The code in this repository has been tested with Python 3.11 but will likely work with other versions of Python and the packages. A detailed list of the packages and versions used can be found in requirements.txt. A brief list is provided below:

• numpy
• pytorch
• pytorch_lightning
• torchrl
• sklearn
• scipy
• pandas
• matplotlib
• seaborn

Case Studies

Three case studies are presented in the paper:

• Beam
• Oscillator
• Population of bridges

Each case study is located in its own subdirectory under cases/. These subdirectories contain:

• All necessary pretrained neural networks, physics-based models, and data files for running the case study.
• Simulator data used to train the Neural Networks, also needed to initialize some of the scalers used in the model.
• An __init__.py file that defines the case study, including different model parameter presets.

Usage Notes

• The main entry points are the scripts in the root directory:
  • 0_single_run.py: Run a single experiment for a selected case study
  • 1_disentanglement_metric.py: Evaluate disentanglement metrics
  • 2_regression_comparison.py: Compare regression performance across cases
• Run the scripts with the working directory set to the root of the folder ./DPI-VAE/.
• Each script contains a case variable to select the case study.
• The 0_single_run and 1_disentanglement_metric scripts additionally contain a preset variable to specify the model preset.
• Multiple model presets can be specified per case study in its __init__.py file.
• Each model preset must specify the model type to use. P corresponds to a model with separate encoders for each modality. S specifies a model with a single encoder.
• The utils/__init__.py file contains global default parameters, which are partially overwritten by the presets defined in each case study.
• The dpivae.py module provides functions for model setup, training, and evaluation, which are used by the main scripts.
• Regularization methods not used in the paper (e.g. adjusting $\alpha$ and $\beta$, constraining the neural network output magnitude and gradient clipping) have not been fully tested and should be used with caution.

References

The simulators used to generate data for the case studies are as follows:

• Beam: We use the beef Python package: https://knutankv.github.io/beef/beef.html.
• Oscillator: Implementation based on the code accompanying the publication: N. Takeishi and A. Kalousis (2021) "Physics-Integrated Variational Autoencoders for Robust and Interpretable Generative Modeling". In Advances in Neural Information Processing Systems 34 (NeurIPS), 2021. Available at the link: https://github.com/n-takeishi/phys-vae.
• Population of bridges: We use the FE model presented in: K. Tatsis and E. Chatzi (2019) - "A numerical benchmark for system identification under operational and environmental variability". 8th IOMAC - International Operational Modal Analysis Conference, Proceedings. Available at the link: https://github.com/ETH-WindMil/benchmarktu1402.