[patnr]

A script that scores PET's update schemes on many twin experiments of a two-phase waterflood, and keeps the results as committed Markdown tables, so that they are easy to refer to and to update as methods are added or changed. It starts with ES-MDA, without and with adaptive localization (autoadaloc) at two thresholds.

What it does

benchmarks/minires/run.py runs each method on many twin experiments of a waterflood: a 1000 x 1000 x 20 m reservoir with an injector on rate and four producers on bottom-hole pressure, over 40 report steps of which the first 16 are history and the remaining 24 the forecast. It does so at two resolutions, coarse (25 x 25, 20 truths) and fine (50 x 50, 8 truths). The unknowns are the log-permeability field and two fluid scalars, sor and vo (#182), and each truth is drawn from the same prior as the ensemble.

Each method is scored on the history misfit, the forecast water cut on the unseen steps, the field (RMSE, correlation, spread) and each scalar (error, spread). The scores are kept as JSON, from which RESULTS.md is generated: one table per case, its header stating the conditions, each cell the mean over the truths with its standard error. Rows are not ranked.

Both cases together take ~4 min on 8 cores.

Results (coarse case, means over 20 truths)

	misfit	RMSE forecast WWCT	logK RMSE	logK corr	logK spread	sor error	sor spread	vo error	vo spread	simulations
prior	91.7	0.221	0.783	0.0201	0.8	0.0314	0.0383	0.676	1.02
ES-MDA (4 steps, α = 4)	1.68	0.0847	0.829	0.205	0.526	0.0349	0.0221	0.602	0.469	250
ES-MDA (4 steps, α = 4), adaptive localization (universal threshold)	29.6	0.146	0.779	0.122	0.773	0.0302	0.0339	0.531	0.894	250
ES-MDA (4 steps, α = 4), adaptive localization (3σ threshold)	17.7	0.123	0.777	0.167	0.752	0.0253	0.0329	0.43	0.773	250

The localization thresholds are the universal one (√(2 ln N) σ) and 3σ, chosen from a comparison of four settings on 3 truths, not from a systematic sweep.

Also

• tests/test_benchmark_minires.py: a slow smoke test, one tiny experiment per method (~5 s), so that the script cannot rot unnoticed. It runs in CI now that CI: install the minires extra, so that the MiniRes tests run #183 installs the minires extra.
• Localization is configured per method; its grid (field) is filled in from the case.
• pyproject.toml: run.py pins the BLAS threads before importing numpy, hence an E402 ignore for that file.

Full suite passes (626 with the minires extra). ruff clean.

🤖 Generated with Claude Code