An Open-source Platform to Facilitate Data-driven FWI Research

Data-driven full waveform inversion (FWI) is a machine-learning-based method thatreconstructs velocity model from seismic data. It learns a direct mapping from seismic data tovelocity model from a large dataset. This method strongly depends on the data amount. Withits increasing popularity, there is an increasing demand for open data and benchmarks. Wepresent an open-source platform http://openfwi-lanl.github.io/ for the study of data-drivenFWI. There are twelve datasets with diverse subsurface structures (flat, curve, etc.) covingmultiple geophysics domains (interface, fault, CO 2 reservoir, etc.) on this website. Thevelocity models are generated with different prior information (mathematical representations,natural images, and geological reservoirs) to cover multiple scenarios of subsurface structuresand the seismic data are simulated with an acoustic wave equation. The benchmarks usingfour deep learning methods, InversionNet (Wu and Lin, 2019), VelocityGAN (Zhang andLin, 2020), UPFWI (Jin et al., 2022), and InversionNet3D (Zeng et al., 2022) are provided asthe baselines for future data-driven FWI research. For those who want to reproduce theseresults, we provide codes and an implementation tutorial. We also manage a Google groupfor people to communicate their research. The platform will be actively maintained and thedatasets are expected to evolve. We hope this platform will benefit the research anddevelopment of data-driven FWI for the community.