DynamisLab/src/SR_analysis/results/runs/article2-plotting-package-20260721/README.md

SR publication plotting package

This package contains five generated publication figures, two replacement tables, and two PPT-ready 16:9 summary pages. All values are derived from frozen canonical artifacts without formula refitting.

• 01_training_case_performance: panel (a) compares PPO and SR absolute closed-loop legacy DTW similarity at 400 control steps for all seven training cases; panel (b) reports each controller's 400-minus-200 change. Here 200 and 400 are run durations in control steps, not Reynolds numbers.
• 02_pointwise_generalization: 200-step SR results at training conditions and sampled unseen interpolation/extrapolation conditions. Each unseen point is one realization. Categories are unconnected because the evidence does not establish a continuous parameter law or statistical robustness.
• 03_steady_rotation_calibration: the retained disturbance-free constant rear-rotation sweep diagnostic.
• 06_example_timeseries_karman_re100: late-window phase portraits and PPO/SR actions over approximately three target cycles. Target indices 96–145 (t_D=38.4–58.0) are used. PPO and SR segments at the same indices are circularly aligned to target sensors_center_uy by mean-centered normalized cross-correlation; deterministic roll lags are PPO=0 and SR=-4 samples. No target actions exist or are plotted.
• 07_flow_field_comparison_karman_re100 (generated by tools/export_flow_comparison.py): target wake, PPO-controlled pinball, and canonical SR-controlled pinball vorticity at one deterministic stable-cycle phase. Over indices [96,146), each trajectory standardizes downstream center-sensor (ux,uy) separately and defines theta=atan2(v_z,u_z). An exhaustive joint search minimizes the two wrapped angular errors plus 0.001 rad/sample times total temporal separation, with a same-direction branch check. Each condition runs once through index 145 while all 50 stable-window fields are held as float32 host-memory candidates (375 MiB total for three 512x1280 buffers); the selected field and phase diagnostic are therefore the exact same CFD sample from the same run. This matches center-sensor limit-cycle phase; it does not assert six-sensor state equality or exact full-field identity. It stores one compressed NPZ, JSON provenance, and PNG/PDF. This is a single snapshot, not an ensemble or uncertainty estimate.
• tables/offline_action_rmse.*: offline SR-vs-PPO next-action RMSE on causally aligned PPO-visited states. This is not closed-loop performance.
• tables/term_deletion.*: actual 40-step closed-loop deletion results, canonical formula identities, same-window parent comparisons, and aggregate mean/min summaries.

The presentation/ directory contains two white-background 16:9 PNG/PDF pages and its source-artifact README. PNG and PDF files share each retained figure stem. phase_alignment.json records the exact Figure 06 selection/alignment contract. manifest.json hashes every package CSV, Markdown, JSON/NPZ metadata or field artifact, and publication figure.