| --- |
| license: mit |
| library_name: pytorch |
| pipeline_tag: image-segmentation |
| tags: |
| - computational-fluid-dynamics |
| - compressible-flow |
| - shock-detection |
| - vortex-detection |
| - multi-label-segmentation |
| - task-preserving-adaptation |
| datasets: |
| - custom |
| metrics: |
| - dice |
| --- |
| |
| # Task-Preserving Shock–Vortex Segmentation (HJ) |
|
|
| This repository contains the three registered checkpoints of the task-preserving |
| harmonized joint (HJ) network used for overlapping shock and vortex-core |
| segmentation in compressible CFD. The model consumes primitive flow fields plus |
| deterministic differential diagnostics and returns independent sigmoid fields; |
| shock and vortex labels are therefore allowed to overlap. |
|
|
| The architecture has a shared encoder with widths `[10, 20, 40, 80]` and |
| task-specific shock and vortex decoders. During corrected-target shock |
| adaptation, the encoder and complete vortex pathway are frozen. Only the shock |
| decoder and projection are updated, giving exact retained-task preservation by |
| construction and direct array replay. |
|
|
| ## Registered results |
|
|
| Means and sample standard deviations over seeds 20260904–20260906: |
|
|
| | Evaluation | Result | |
| |---|---:| |
| | Airfoil shock Dice, corrected weak reference | 0.936 ± 0.004 | |
| | Cylinder shock Dice, corrected weak reference | 0.939 ± 0.024 | |
| | Retained airfoil core Dice | 0.827 ± 0.048 | |
| | Retained cylinder core Dice | 0.228 ± 0.020 | |
| | Independent analytic-vortex disk Dice | 0.672 ± 0.086 | |
| | Independent analytic centre recovery | 0.889 ± 0.192 | |
| | Continued-SU2 finite-ray coverage | 1.000 ± 0.000 | |
| | Continued-SU2 conditional normal MAE | 0.00284c ± 0.00092c | |
|
|
| Core probability drift is exactly zero for all three seeds over all 129 |
| seed–field evaluations. These results do **not** establish human-ground-truth |
| accuracy, universal geometry transfer, or complete-front validation. The CFD |
| Dice references are physics-derived weak references; the analytic vortex and |
| finite SU2 rays are distinct bounded controls. |
|
|
| ## Files |
|
|
| - `checkpoint_seed2026090X.pt`: task-preserving HJ checkpoint. |
| - `thresholds_seed2026090X.json`: validation-selected operating point. |
| - `aggregate.json`: complete three-seed aggregate used by the manuscript. |
| - `MODEL_CARD_LIMITATIONS.md`: intended use and limitations. |
|
|
| ## Code, media and demo |
|
|
| The Space displays the extended multistructure pipeline used in Movies S9/S10: |
| the learned local shock-front model with fixed front extraction, retained |
| core/wake branches and the PM-v4 expanding-flow region. This pipeline differs |
| from the three registered HJ control checkpoints in this repository. The |
| metrics above describe those HJ controls, not the extended movie pipeline. |
| The Space includes the figure provenance and links to the movie release. |
|
|
| - [Source code and reproducibility material](https://github.com/Ehsan-Roohi/ShockVortexML) |
| - [Original-quality supplementary movies](https://github.com/Ehsan-Roohi/ShockVortexML/releases/tag/movies-localfront-v2-20260909) |
| - [Author’s YouTube channel](https://www.youtube.com/@ehsan_roohi) |
| - [Interactive Space](https://huggingface.co/spaces/ehsanroohi/ShockVortexML-Demo) |
|
|
| ## License and citation |
|
|
| Released under the MIT License. The manuscript is under review/preparation; |
| please use the repository citation metadata until a DOI is assigned. |
|
|
