docs: add frequency-selection FWI notebook (streamer, multiscale)

[GeophyAI]

Adds notebook 31: frequency-selection source encoding (Tromp & Bachmann, 2019) written out in ~60 lines of torch on a plain PropTorch solver. No dependency on sweep-tasks.

Setup

• Marmousi, towed streamer: 129 shots every 100 m, 4 km cable, all shots in one encoded forward.

• Multiscale, with each band on its own grid and time step (Courant 0.38):

  band	grid	dt
  1.5–3 Hz	100 m	8 ms
  2.5–5 Hz	50 m	4 ms
  4–10 Hz	25 m	2 ms

• The observed data are transient Ricker gathers, DTFT'd onto each band's comb. The inversion uses no wavelet.

Checks in the notebook

• Two-window steady-state check: ≤3.3e-3 at a 7 s ring-up.
• GCN misfit on the true model ≈1e-5 per shot.
• The unfired spare bin sits orders of magnitude below the fired bins (zero crosstalk).
• RMSE goes 385 → 378 → 345 → 299 m/s across the three bands.

Section 7: speedup per iteration
Compared against conventional FWI that also uses all 129 shots, on the same grid:

band	theory	measured (RTX 6000 Ada, conventional at its fastest batch size)
1.5–3 Hz	13.7×	0.5×
2.5–5 Hz	21.7×	2.9×
4–10 Hz	44.2×	18.1×

The notebook explains the gap: one encoded wavefield cannot fill the GPU on small grids, while the conventional run fills it by batching shots.

Also changed

• mkdocs nav entry and examples-page card.
• Gallery thumbnail.

Executed end-to-end on the rebased tree (dacf1f61), about 3 min on one Ada.

🤖 Generated with Claude Code