Skip to content

cluster_regime_narrative.md

Latest commit

 

History

History
92 lines (81 loc) · 4.76 KB

cluster_regime_narrative.md

File metadata and controls

92 lines (81 loc) · 4.76 KB
Raw

Cluster package: point-source-default narrative + extended-source follow-up feature

Type: docs Target: autolens_workspace Repos:

  • autolens_workspace
  • autolens_workspace_test Difficulty: medium Autonomy: supervised Priority: high Status: draft Epic: cluster-strong-lensing Phase: 10 Parent: draft/docs/autolens/split_lensing_regimes.md Filed: 2026-07-25 (backfilled from git)

Align the scripts/cluster/ package of @autolens_workspace with the three-regime design (see parent plan). The mass framework (host halo(s) + truncated members + scaling relations) is intentionally SHARED with the group regime — the cluster regime is distinguished by the observational setting and therefore the SOURCE MODELLING STRATEGY: dozens–hundreds of members, many multiply-imaged sources across a wide redshift range, so the default workflow is multiple-image positions / point-source constraints with individual source redshifts, jointly optimizing one cluster mass model (multi-plane). Extended source reconstruction is a specialised follow-up analysis of individual systems, NOT the default.

Much of this package already exists (real Abell 2744 start_here on the Bergamini et al. 2023 gold sample, dPIE + scaling relations, CSV API, LensTool interop, mass_parameterizations guide). This task is narrative alignment + gap-filling, not a rebuild.

Changes

  • README.md + start_here.py prose: state the regime-ladder design explicitly — same mass framework as group/, different source strategy; all clusters are multi-galaxy systems but not vice versa; link down the ladder to group/ and multi_galaxy/.
  • New features/extended_source/ follow-up example: take one system from the cluster fit (e.g. one A2744 arc) and do a targeted extended-source reconstruction (imaging + pixelized source) with the cluster mass model as the starting point — framed explicitly as the specialised follow-up, and as the bridge back to the group/galaxy-scale source machinery. Note the foreground lens light is NOT modelled in the default cluster workflow (a deliberate divergence from @autogalaxy_workspace's cluster package, which is about the galaxies' light); an autolens lens-light cluster feature is future work, out of scope here.
  • Conventions sweep — the guides (mass_parameterizations.py, mass_parameterizations_pyautolens.py) and the dPIE library docstrings were corrected on 2026-07-25 (Bergamini et al. 2019 tied exponents with gamma=0.2, vanishing unscaled member cores, r_cut_ref ~5", sigma_LT vs sigma_0 attribution per the H. Ding derivation note). STILL TO SWEEP here: start_here.py (scaling_radius_exponent=0.5, r_core scaled with L, r_cut_ref 15.8"), simulator.py (same dated truths) and modeling.py (refs fixed at those truths). The simulator truths and the bundled cluster/simple dataset must change TOGETHER — re-run the simulator and commit the regenerated dataset in the same PR, then re-validate modeling and start_here end-to-end (this is why the sweep was deferred to this child task rather than done alongside the guides).
  • gNFW guidance: ensure the "beyond the LensTool default" prose (dPIE host → (G)NFW host) is present and linked from start_here, per the expert feedback recorded in the parent plan.
  • Ground the narrative in the parent plan's cluster literature section (HFF/CLASH/JWST-era benchmarks, model-comparison projects) with citations. Specifics from the research: state that start_here uses the PRE-JWST spectroscopic gold subset of Bergamini et al. 2023 by design (JWST-era models of A2744 now use ~135–150 images); name AS1063 as the "simplest relaxed cluster" counterpoint to merging A2744; candidate future feature systems — MACS J0416 (largest spec sample, scaling-up), SMACS J0723 (mid-size, JWST-iconic), MACS J1149/SN Refsdal (time-delay cosmography), SDSS J1004+4112 (cluster-lensed quasar bridge from point_source users).

Conventions-sweep validation note (2026-07-25)

The sweep of simulator.py/modeling.py/start_here.py + regenerated cluster/simple dataset landed on the task branch. modeling.py validated green end-to-end (PYAUTO_TEST_MODE=2) against the regenerated data. start_here.py (real a2744, 188 members, multi-plane) did NOT complete a 45-minute bypass-mode run on the dev container's CPU — the JAX compile of the point solver dominates and is unchanged by the sweep (constants only; 2 fewer free parameters). It is not smoke-gated. First GPU session should run it once to confirm end-to-end (expected ~10 min).

autolens_workspace_test

Add/extend a cluster extended-source follow-up integration script.

Acceptance

  • Smoke suite green; notebooks + navigator regenerated.
  • A new user reading cluster/README understands why cluster examples fit positions not pixels, and where the extended-source path lives.