BayesFrag is an open-source Python library to perform Bayesian parameter estimation of empirical seismic fragility models. The methodology is presented in
Bodenmann L., Baker J.W. , Stojadinović B. (2023): "Accounting for ground motion uncertainty in empirical seismic fragility modeling", Engineering Archive [preprint], doi:10.31224/3336.
Seismic fragility functions provide a relationship between damage and a ground motion intensity measure (IM). Empirical fragility functions are estimated from damage survey data collected after past earthquakes. This is challenging, because the IM values that caused the observed damage are uncertain. BayesFrag computes the joint posterior distribution of fragility function parameters and IM values at the sites of the surveyed buildings.
To avoid any additional dependency on a specific library of ground motion models (GMMs), the GMM estimates are computed prior and separate to the fragility parameter estimation.
To allow for a smooth start, we offer three tutorials that can be opened on a hosted Jupyter notebook service (e.g., Google Colab).
- The first tutorial reproduces the results of the one-dimensional example from the manuscript and explains the implementation of the proposed estimation approach:
- The second tutorial shows how the Bayesian approach can be applied to a realistic damage survey data set. The notebook can easily be modified and extended such that analysts can apply BayesFrag to their data sets of interest:
- The third tutorial explains the computation of the GMM estimates, which is done prior and separate to the fragility parameter estimation:
The folder bayesfrag contains the main codebase of the BayesFrag package.
The folder data contains the data used in the two tutorials.
The folder results contains the results illustrated in the manuscript.
The notebook figures.ipynb reproduces the figures shown in the manuscript using utility functions from utils_plotting.py.
The script main_aquila_casestudy.py performs all computations - from parameter estimation to post-processing - to reproduce the results presented in the paper (and stored in the folder results). This script is computationally expensive and causes high memory demand. The computations were performed on a NVIDIA V100 tensor core GPU with 25 GB of RAM and we do not recommend to run this script on a standard personal computer.
BayesFrag can be installed as
pip install bayesfrag
It is good practice to do the installation in a new virtual environment, for example with venv or with miniconda. BayesFrag was tested with Python >= v3.10.
BayesFrag is an open-source software. If you are interested in contributing to the project or have any questions, comments, or suggestions, please contact Lukas Bodenmann at bodenmann (at) ibk.baug.ethz.ch.
We gratefully acknowledge support from the ETH Risk Center ("DynaRisk", Grant Nr. 395 2018-FE-213) and from the Chair of Earthquake Engineering and Structural Dynamics at ETH Zurich.
BayesFrag is released under the BSD-3-Clause license.