OpenFWI: Seismic Full Waveform Inversion

Full waveform inversion reconstructs high-resolution subsurface velocity maps from seismic records, governed by the acoustic/elastic wave equation and formulated as a non-convex optimization. OpenFWI provides 2.1 TB of synthetic multi-structural datasets spanning interfaces, faults and CO2 reservoirs, plus benchmarks of four deep learning methods (InversionNet, VelocityGAN, UPFWI, InversionNet3D).

Key findings: Data-driven FWI generalizes poorly across domains, and the degradation correlates with subsurface structural complexity - making OpenFWI a useful stress test rather than a solved benchmark.