2026-10-12T10:56:48Z https://api.figshare.com/v2/oai

oai:figshare.com:article/33993865 2026-09-25T09:52:34Z category_29173 category_26581 item_type_3 month_year_09_2026

HiLiftAeroML: score reproduction and paper checkpoints neil ashton (25108192) Adam Clark (7838714) Konrad Goc (25108246) Liam Heidt (25108247) Christopher Ivey (25108250) Rahul Agrawal (193213) Sanjeeb T Bose (25108253) Corey Adams (10925682) Peter Sharpe (25108256) Daniel Leibovici (25108259) Semit Akkurt (25108268) Sheel Nidhan (25108269) Rishi Ranade (25108272) Jean Kossaifi (25108282) Machine learning not elsewhere classified Aerodynamics (excl. hypersonic aerodynamics) HiLiftAeroML CFD GeoTransolver Transolver reproducibility high-lift aerodynamics PhysicsNeMo <p dir="ltr">Companion reproducibility artifacts for HiLiftAeroML: A High-Fidelity Computational Fluid Dynamics Dataset for High-Lift Aircraft Aerodynamics. Version 1.0.0.</p><p dir="ltr"><br></p><p dir="ltr">The CPU package replays retained score reductions for 23 model-split records and 8,142 case evaluations using only the Python standard library. The checkpoint package contains all 46 matching GeoTransolver/Transolver model exports, 23 surface normalizers, configurations, constructor metadata and the HiLift training recipe. Both packages include manifests and SHA-256 checksums.</p><p dir="ltr"><br></p><p dir="ltr">The paper uses a preliminary version of the FluidsBench evaluation metrics, with the scoring code frozen to revision e19809a194ee1cbc4b16d4212574f8f4d921cc40 for this study. These metrics may differ from the final FluidsBench release. The included FluidsBench source is limited to the scoring module used by the replay and its license.</p><p dir="ltr"><br></p><p dir="ltr">Verified scope: reduction of retained per-case field statistics, coefficient pairs and profile R2 into the checked aggregate metrics. Fresh native inference, force integration, profile scoring from evaluator truth, and training are not rerun by this release. PhysicsNeMo framework versions, runtime records and training settings are documented.</p><p dir="ltr"><br></p><p dir="ltr">The checkpoint weights are licensed under CC BY 4.0. Existing source-code licenses, including Apache-2.0, remain in force for their corresponding files; see LICENSING.md and the retained source notices. The underlying CFD dataset is available separately at https://huggingface.co/datasets/nvidia/HiLiftAeroML.</p><p dir="ltr"><br></p><p dir="ltr">Use hilift-score-reproduction-arxiv.zip for CPU score replay and hilift-checkpoints-arxiv.zip for the corresponding model exports. Verify the downloads with ARCHIVE_SHA256SUMS.txt.</p> 2026-09-25T09:52:34Z Dataset Dataset 10.6084/m9.figshare.33993865.v1 https://figshare.com/articles/dataset/HiLiftAeroML_score_reproduction_and_paper_checkpoints/33993865 CC BY 4.0