Forward modelling

task_type: forward propagates a wavelet through a model and records it at the receivers. Three examples, smallest last: 01 is the Marmousi record the inversion pages invert, 09 is the smallest task there is, and 10 places sources and receivers by hand.

01 — a Marmousi shot record

sweep-tasks run examples/synthetic/01_forward_marmousi.yaml

Propagates an 8 Hz Ricker through vp_true with 114 shots at 150 m and a 1361-channel fixed spread, recording 10 s at dt = 1 ms. It produces output/record.npy with shape (114, 10000, 1361, 1) plus sources.npy and receivers.npy.

Reference (RTX 6000 Ada, seed 0): 34 s.

The first arrival starts at 2 s rather than 0: the wavelet carries 2 s of lead-in (wavelet.delay: 2.0), which 03 needs so that low-passing the Ricker to 0.5–2 Hz does not clip its left lobe. 01–03 share one wavelet so their results are directly comparable.

Two properties of this acquisition drive the choices in the FWI examples:

• obs dominant frequency is 7.10 Hz, not the source's 8 Hz — propagation and geometric spreading shift the peak down. Half a period is 70 ms.
• the latest first break lands at 8.59 s of the 10 s record. A 7 s record (the obvious first guess) cut the far offsets off right after their first arrival: 31.7 % of the edge shots' traces kept under 1 s of coda. At 10 s that figure is 0 %.

09 — the smallest task there is

sweep-tasks run examples/synthetic/09_forward_constant_box.yaml

One shot through a constant-velocity box. Its point is the YAML, not the result: it annotates every field of a task spec — what it does, what the alternatives are, what the default means — and the other examples assume that vocabulary. It needs no dataset and no GPU (it runs on eager), and produces output/record.npy with shape (1, 500, 172, 1).

A constant model has no reflections, so the record is the direct arrival and nothing else. That is what makes it a good first test: anything other than one clean hyperbola means the setup, not the geology, is wrong.

10 — hand-placed sources and receivers

sweep-tasks run examples/synthetic/10_forward_explicit_geometry.yaml

The same box as 09, with geometry.kind: explicit — literal coordinate lists instead of a stride rule. Three shots, seven receivers, record shape (3, 500, 7, 1). Only the geometry: block differs from 09.

The four geometry kinds, and when each earns its place:

geometry.kind	what it is	use it for
line	stride rule, 2-D only	regular 2-D surveys — the default
grid	stride rule on x and y, 3-D only (13)	regular 3-D patches
explicit	literal lists, as here	irregular or small layouts — a gap in the spread, one odd node
from_file	the same arrays as .npy	layouts a script generated

from_segy_headers, from_segy_index and from_plan read real acquisition geometry out of field data; the Viking example uses those.