Skip to content

FWI, single-scale

sweep-tasks run examples/synthetic/02_fwi_marmousi_single.yaml

FWI from vp_smooth, the low-pass-filtered true model. It keeps the right depth trend, so a single broadband band converges without cycle-skipping. Read this one first among the inversions: it shows the FWI task shape, obs.synthetic_from (the observed data are synthesised from vp_true on the fly, with the same grid, geometry and wavelet as 01), the boundary-saving adjoint and the per-epoch QC — without the multiscale machinery on top.

Reference (RTX 6000 Ada, seed 0), 200 epochs, 29.6 min:
misfit 1.592e-02 → 1.564e-04 (min, epoch 115) → 2.575e-04
RMSE vs true 357.3 (init) → 270.6
perturbation correlation r 0.700

inverted vp

misfit

The runner's obs/syn QC panel — interleaved gathers, the acquisition map, and a receiver-averaged amplitude spectrum where obs and syn sit on top of each other:

obs vs syn

Where the 30 minutes go

200 epochs × 16 shots is 3200 shot gradients, each a 10 s record at dt = 1 ms (10,000 steps). With strategy: boundary the backward pass re-propagates the forward from the saved boundary, so each gradient is about three wavefield sweeps. Run it on several GPUs to split the shots:

sweep-tasks run examples/synthetic/02_fwi_marmousi_single.yaml --nproc-per-node 4

batchsize is a fraction, not a count

It is how many shots are drawn at random per optimizer step, so what matters is batchsize / nshots. Keep it near ~15 %. Measured here: 114 shots with batchsize: 8 (7 %) gives RMSE 297, and batchsize: 16 (14 %) gives 271 — denser shots only pay off if the batch grows with them.