FWIPrePostAI - Shear Wave Velocity Prediction with Data-Driven Full Waveform Inversion (FWI) Models.
Sanish Bhochhibhoya and Joseph P. Vantassel
FWIPrePostAI is a Python-based toolkit designed to predict shear wave velocity profiles from waveform datasets using pre-trained artificial intelligence (AI) models. The repository includes tools for working with both simulated and real-field data, and features a pre-trained model developed by Vantassel and Bhochhibhoya (2025).
This work facilitates rapid and data-driven full waveform inversion (FWI) for geotechnical applications, lowering the technical barrier for non-expert users while maintaining robust predictive performance.
If you use FWIPrePostAI in your research or consulting, we ask you please cite the following:
Vantassel, J. P. and Bhochhibhoya, S. (2025). “Toward More-Robust, AI-enabled Subsurface Seismic Imaging for Geotechnical Applications.” Computers and Geotechnics. [In-Review]
-
Ensure Python is Installed.
Verify that Python version 3.10 or later is installed on your system. If not, please follow the detailed installation instructions provided here. -
Download the Repository.
Obtain a local copy of the FWIPrePostAI repository, either by downloading the repository as a .zip file and extracting its contents, or cloning the repository via Git. -
Install Required Python Packages.
Navigate to the repository directory and install the necessary Python dependencies using the following command:
python -m pip install -r requirements.txt
For users unfamiliar with pip, a comprehensive tutorial is available here. -
Download the
dataFolder.
Download thedatadirectory from the provided Google Drive Link. Ensure that the downloadeddata,fwiprepostai, and.ipynb filesare all placed in the same parent directory. More details on the contents ofdatafolder are provided in a later section.
-
Launch the Notebooks.
Open eitherData_Driven_FWI_Prediction_field_data.ipynborData_Driven_FWI_Prediction_simulated.ipynbto get started with the prediction models.Data_Driven_FWI_Prediction_field_data.ipynb- Notebook to perform a prediction on field data using the Vantassel and Bhochhibhoya (2025) model.Data_Driven_FWI_Prediction_simulated.ipynb- Notebook to perform a prediction on the simulated (testing) dataset using the Vantassel and Bhochhibhoya (2025) model.
-
Explore the Notebooks.
Experiment with the notebooks using the different data provided. Once comfortable, try applying the Vantassel and Bhochhibhoya (2025) model to field or simulated data of your own.
The data directory includes the following files and subdirectories essential for running the notebooks and reproducing results:
model_weights/VantasselAndBhochhibhoya2025.weights.h5contains the pre-trained model weights corresponding to the Vantassel and Bhochhibhoya (2025) architecture. This file is compatible with the model class defined infwiprepostai/data_driven_fwi_models.py.waveforms\generated_data\samples_500_116.h5contains a subset (500 samples) out of 10,000 sample testing dataset that used during the testing of the Vantassel and Bhochhibhoya (2025) model.
Note: Samples with indices 269, 87, 252 were used in Figure 4 of the Vantassel and Bhochhibhoya (2025) paper.waveforms\real_field_data_numpycontains two NumPy arrays that represents the raw (drillfield_June28_raw.npy) and the field-corrected (drillfield_June28_lisousi.npy) data collected on the Virginia Tech Drillfield. \ Applying the correction procedure proposed by Forbriger et al. (2014) transforms the raw data to the lisousi data.
Note: When using real-field data, you must apply the correction procedure described in Forbriger et al. (2014). We present both the raw and corrected data to allow others to compare their correction process to the one we have used for developing the AI model.waveforms\real_field_data_geodecontains ten raw real-field waveform dataset obtained during field experiments at the Drillfield, Blacksburg, VA. Stacking these ten datasets reproduces drillfield_June28_raw.npy, as mentioned above.