[lispandfound]

Summary

Adds depth-dependent ("hybrid") slip correlation lengths, after Graves & Pitarka's SCEC 2026 poster GM-218 and genslip v5.6.3 (SCEC BBP dev branch). Shallow slip can use long correlation lengths (Suzuki et al. 2022) and deep slip short ones (Mai & Beroza 2002), crossfaded between two depths. It also supports any number of families.

GP blends two fields linearly, so the latent variance drops in the transition, to 0.845 at Mw 7 with shared noise and 0.5 with independent noise. This PR blends in quadrature instead:

• Every family filters one shared white noise, embedded on a common padded grid.
• Families are crossfaded at equal power: loading sin(πw/2), normalised by the exact cross-correlations between families. The latent field is therefore standard normal at every cell, and the marginal transform stays exact.

The structure was settled through two architecture reviews, a devil's-advocate pass, /simplify and /code-review. Backwards compatibility was explicitly not a constraint.

Changes

• sampling/field.py: one Sampler type for any number of covariances. sampler(grid, covariances, weights=None) embeds them at the widest padded shape and draws all of them from one seed. The stationary case is one covariance. The lru_cache on sampler is gone.
• rupture/models.py:
  • NoiseFamily(relation, hurst, weight), with Noise = tuple[NoiseFamily, ...] and STATIONARY as the default noise.
  • embed(noise, geometry, magnitude, marginal) pre-corrects each family for the marginal, reads the weights at cell centres, and caches the result.
  • DrawnSlip.noise, GaussianRake.noise and CorrelatedPerturbation.noise let each model choose its noise. Making rake hybrid is a config change. By default rake stays stationary and the perturbation uses slip's noise, which matches genslip's slip and rake. genslip keeps the perturbation's own part stationary; set perturbation.noise to STATIONARY to reproduce that.
  • SlipField.sampler replaces latent_covariance. Rise time and the onset perturbation draw their independent part from slip's built sampler, so nothing is embedded twice.
• rupture/source.py:
  • SegmentSource is now a moment and a rake, with a magnitude property.
  • CorrelationRelation gains strike_saturation and dip_saturation, for SEA22's min(Mw, 6.3).
  • New SUZUKI_2022 relation.
• config.py: [[slip.noise]], [[rake.noise]] and [[perturbation.noise]] replace [source.correlation] and [segments.<name>.correlation]. A segment's noise list replaces the global one whole. Profile depth/value lists are now tuples, so profiles are hashable.
• Examples and README are updated to match.

Breaking changes (intentional)

• Configs with [source.correlation] no longer load. The lengths now live on each model's noise.
• SegmentSource.covariance, SlipField.latent_covariance, segment_source's correlation arguments, and sampler.cache_clear() are removed. The cache is now embed.cache_clear().
• Draws differ slightly from before, because stationary fields are now normalised exactly for clipped eigenvalues.

Example

[[slip.noise]]   # shallow: Suzuki et al. (2022)
relation = { strike_offset = 1.67, dip_exponent = 0.5, dip_offset = 1.69, dip_saturation = 6.3 }
weight = { type = "ramp", centre_km = 6.5, half_width_km = 1.5, shallow = 1.0, deep = 0.0 }

[[slip.noise]]   # deep: Mai & Beroza (2002), the default relation
weight = { type = "ramp", centre_km = 6.5, half_width_km = 1.5, shallow = 0.0, deep = 1.0 }

Testing

• pytest tests: 247 passed, 1 skipped.
• Doctests, ruff and ty pass; ty only reports the existing unresolved rerun import.
• pytest benchmarks/test_kernels.py --benchmark-disable passes.
• New tests/test_blend.py covers:
  • saturation;
  • the shared padded shape;
  • a cell under one family equalling that family's own draw;
  • empirical unit variance and independent real and imaginary parts through the crossfade;
  • weight validation;
  • slip versus rake noise ownership;
  • the perturbation's own noise;
  • config round trip and segment overrides.

Not in this PR

• GP's depth-dependent along-strike side taper (taper_slip_all_r_hc).
• A Rust kernel that generates the white noise once for all K transforms.
• Draws with several families cost K transforms on the largest family's padded grid. At Mw 7 with SEA22 that is about 23× the stationary MB02 cost.

---

Generated by Claude Code

[github-actions]

Benchmarks

31f9a28 against base cafe752, on AMD EPYC 9V45 96-Core Processor. Times are medians over the rounds. The median change is 2.3%; changes past 5% are in bold.

Kernels

kernel	base	head	change	head IQR	rounds	padded cells
embed[Mw6.5]	28.2 ms	27.7 ms	-1.8%	260 µs	35	448,000 (+0.0%)
circulant_draw[Mw6.5]	2.61 ms	2.68 ms	+2.8%	250 µs	290
embed[Mw7.5]	88.3 ms	88.2 ms	-0.1%	1.03 ms	9	1,204,224 (+0.0%)
circulant_draw[Mw7.5]	7.48 ms	6.55 ms	-12.4%	1.27 ms	150
eikonal_solve	950 ms	937 ms	-1.4%	13.2 ms	5
synthesise_pulses[impulse]	373 µs	364 µs	-2.5%	17.5 µs	523
synthesise_pulses[liu]	430 ms	414 ms	-3.8%	5.34 ms	5

End to end

Each round is the command line in a fresh interpreter, start-up included. Peak RSS is the largest the child process reached over the rounds.

example	base time	head time	change	base peak RSS	head peak RSS	change
alpine_hope	14.1 s	13.2 s	-6.2%	445 MiB	371 MiB	-16.6%
beavan	3.63 s	3.43 s	-5.6%	440 MiB	396 MiB	-10.0%
colombia	2.83 s	2.79 s	-1.4%	1,124 MiB	1,128 MiB	+0.4%
crustal	827 ms	810 ms	-2.0%	188 MiB	190 MiB	+0.6%
landers	3.27 s	3.08 s	-5.7%	797 MiB	806 MiB	+1.1%
northridge	1.07 s	1.05 s	-1.4%	349 MiB	349 MiB	-0.0%
two_faults	1.28 s	1.3 s	+1.7%	261 MiB	266 MiB	+1.9%
two_faults_segments	995 ms	961 ms	-3.4%	199 MiB	201 MiB	+1.0%
venezuela	6.23 s	6.17 s	-0.9%	2,412 MiB	2,345 MiB	-2.8%