Full-Waveform Inversion (FWI)

---

Full-waveform inversion updates the velocity model by fitting the complete recorded wavefield, not only travel times. BFI has built its own 2D acoustic FWI engine on GPU and applied it to reflection data from the public Mobil Viking Graben Line 12 dataset in the North Sea, starting from a travel-time tomography model.

48 shots were inverted. Every result is measured on the 47 shots in between that the inversion never used, each 125 m from an inverted shot, over the full 4-28 Hz band. This shows the model predicts new data rather than memorising the shots it was fitted to.

On the 47 held-out shots: data misfit fell to 0.36 of the starting model; correlation with the recorded data rose from 0.38 to 0.78; the systematic travel-time lag dropped from +8.2 ms to +0.3 ms; every held-out shot improved in every frequency band; and held-out and inverted shots differ by only 2.3 percentage points, so the model generalises.

Direct and diving waves are muted, and surface multiples are kept and modelled with a free surface. The water layer of the tomography model is rebuilt at the picked seabed. The source wavelet comes from linear source inversion with ghost correction. The inversion runs in three frequency bands (up to 10, 18 and 28 Hz) with shot-footprint filtering, and every update is saved.

Two published velocity logs on the line were each digitised twice, independently, and compared with the models at matched resolution. Down to about 2 km, vertical time through the final model agrees with both logs to 2-3%, with the same sign as the travel-time lag measured on the shots. Well logs, migrated images and held-out shots agree three ways.

The full 34-iteration inversion ran in 15.9 hours on a single laptop GPU. Our fused-kernel version of the engine is 5.4 times faster than the reference code and gives bit-identical results, so FWI fits within a normal project schedule.