OpenRadioss Dataset Generation#

End-to-end recipe for generating parameterized OpenRadioss simulation datasets and converting the output to the LS-DYNA-style d3plot layout expected by PhysicsNeMo-Curator — the upstream step for training structural dynamics surrogates in examples/structural_mechanics/crash/ (for bumper beam) and examples/structural_mechanics/drop_test (for drop test).

Two flows are supported out of the box:

Flow

Base case

Varied parameters

Default DoE size

bumper_beam

Front-impact bumper beam

geometry scale, shell thickness, impact velocity, rigid-wall diameter/origin

135 runs

drop_test

Cell-phone drop onto a plane

per-material Young’s modulus (E), rigid-wall plane orientation (rx, ry, rz)

192 runs

Both flows share the same runner, d3plot-renaming step, and summary helpers; only the parameter-mutation logic differs.

Datasets#

Each flow is a Design-of-Experiments sweep: the generator takes the Cartesian product of the configured axes, writes one case folder per combination, and applies the mutations directly to the _0000.rad starter deck. The defaults below live in DEFAULT_EXPERIMENT_SETUP near the bottom of each flow’s generate_dataset.py — edit or replace that dict to resize or reshape the sweep.

Bumper beam#

Front-impact bumper-beam setup driven by an impactor on a rigid wall. The sweep targets the crash-response envelope by perturbing geometry and loading: bumper shape, wall-thickness mass distribution, impact speed, and the impactor’s relative location on the beam.

Default: 5 × 3 × 3 × 1 × 3 = 135 runs.

#

Parameter

Deck target

Default values

Count

1

Geometry scale ``(sx, sy, sz)`` — per-axis multiplier applied to every node

/NODE block

(1,1,1), (1,0.5,1), (1,1,0.5), (1,2,1), (1,1,2)

5

2

Initial velocity ``(vx, vy, vz)`` of the impactor (mm/ms)

/INIVEL line 2

(-5,0,0), (-3,0,0), (-7,0,0)

3

3

Shell thickness multiplier

/PROP/SHELL line 4, col 3

1.0, 0.7, 1.3

3

4

Rigid-wall diameter (mm)

/RWALL line 3, col 3

254.0

1

5

Rigid-wall origin ``(x, y, z)`` — M is replaced; M1 shifts by the same delta so the normal is preserved

/RWALL lines 4 + 5

(-170,0,0), (-170,120,0), (-170,240,0)

3

Global-feature columns emitted for training: geo_scale_{x,y,z}, velocity_{x,y,z}, thickness_scale, rwall_diameter, rwall_origin_{x,y,z}.

Drop test#

Cell-phone drop onto a plane. The sweep targets material stiffness × impact pose — perturbing Young’s modulus for every deformable component while tilting the contact plane — to cover the stiffness/orientation interactions that drive peak stress and deflection.

Default: 2⁵ × 6 = 192 runs.

#

Parameter

Deck target

Default values

Combinations

1

Young’s modulus scale per material — scales the E value on the line after the # E Nu comment in each /MAT/ELAST or /MAT/PLAS_TAB block

mat 1 (polymer), mat 4 (battery), mat 5 (glass), mat 8 (PCB), mat 9 (composites)

each mat ∈ {0.8, 1.2} (±20%)

2⁵ = 32

2

Rigid-wall plane orientation ``(rx, ry, rz)`` — rotates the M→M1 normal about M by the given XYZ Euler angles (degrees); M1 is rewritten

/RWALL line 5

(0,0,0), (±10,0,0), (0,±10,0), (0,0,10)

6

Global-feature columns emitted for training: e_scale_mat{1,4,5,8,9}, rwall_orientation_{rx,ry,rz}.

Directory layout#

openradioss_dataset_gen/
├── README.md
├── common/
│   ├── radioss_runner.py        # Starter / Engine / VTK / D3PLOT batch runner
│   ├── rename_d3plot.py         # Rename <BASE>.d3plot* -> d3plot*
│   └── summary_utils.py         # summary.json + summary.csv writer
├── bumper_beam/
│   ├── generate_dataset.py      # DoE generator (see DEFAULT_EXPERIMENT_SETUP)
│   ├── run_simulations.py       # Thin wrapper over common.radioss_runner
│   ├── restructure_global_features.py   # summary.json -> global_features.json
│   └── templates/               # place Bumper_Beam_AP_meshed_0000.rad / _0001.rad here
└── drop_test/
    ├── generate_dataset.py
    ├── run_simulations.py
    ├── restructure_global_features.py
    └── templates/               # place Cell_Phone_Drop_0000.rad / _0001.rad here

Prerequisites#

  • OpenRadioss built with the GFortran-Linux executables (starter_linux64_gf, engine_linux64_gf, anim_to_vtk_linux64_gf). Point the runner at your build via the OPENRADIOSS_ROOT env var.

  • ``vortex_radioss`` for the anim -> d3plot conversion step:

    pip install vortex-radioss

Base case inputs#

The .rad starter and engine files are not shipped in the recipe because the bumper deck is ~1.8 MB and the drop-test deck is ~42 MB. Fetch them from the upstream OpenRadioss sources and place them under the corresponding templates/ directory:

Flow

Files to place under templates/

Source

bumper_beam

Bumper_Beam_AP_meshed_0000.rad, Bumper_Beam_AP_meshed_0001.rad

https://openradioss.atlassian.net/wiki/spaces/OPENRADIOSS/pages/11075585/Bumper+Beam

drop_test

Cell_Phone_Drop_0000.rad, Cell_Phone_Drop_0001.rad

OpenRadioss sample models (Cell Phone Drop example)

Workflow (same for both flows)#

All commands below assume you are in the flow directory (bumper_beam/ or drop_test/).

1. Generate the parameterised dataset#

python generate_dataset.py

Creates: - dataset/run1/, dataset/run2/, … each containing a mutated _0000.rad starter, a copy of the _0001.rad engine deck, and a per-run <run>.json metadata file. - dataset/summary.json — hierarchical log of every run. - dataset/summary.csv — flat, ML-loader-friendly view.

To customise the DoE, edit DEFAULT_EXPERIMENT_SETUP near the bottom of generate_dataset.py, or import generate_dataset(...) from another driver script.

2. Run the simulations#

export OPENRADIOSS_ROOT=/path/to/OpenRadioss   # required
export MAX_PARALLEL_JOBS=4                     # optional
export OMP_NUM_THREADS=8                       # optional
python run_simulations.py

For each run*/ folder, the runner executes: 1. starter_linux64_gf -i <BASE>_0000.rad -nt $OMP_NUM_THREADS (log: starter.log) 2. engine_linux64_gf  -i <BASE>_0001.rad (log: engine.log) 3. anim_to_vtk_linux64_gf <BASE>A### > <BASE>A###.vtk per animation frame 4. vortex_radioss.animtod3plot.Anim_to_D3plot.readAndConvert(<BASE>) — emits <BASE>.d3plot, <BASE>.d3plot01, … (log: d3plot_conv.log)

Total wall time scales with MAX_PARALLEL_JOBS * OMP_NUM_THREADS; do not oversubscribe CPU cores.

Set DEBUG_MODE=1 to run only the first case.

3. Rename d3plot files to the LS-DYNA convention#

python ../common/rename_d3plot.py \
    --dataset-dir ./dataset \
    --base-name Bumper_Beam_AP_meshed      # or Cell_Phone_Drop

PhysicsNeMo-Curator expects the plain d3plot, d3plot01, … layout; the runner produces <BASE>.d3plot*. This step renames them in place.

4. Restructure global features for the datapipe#

The recipes’ datapipe expects a global_features.json keyed by run ID (see examples/structural_mechanics/crash/README.md, Global features section). Convert dataset/summary.json into that format:

python restructure_global_features.py                 # writes ./global_features.json

Bumper-beam keys: geo_scale_{x,y,z}, velocity_{x,y,z}, thickness_scale, rwall_diameter, rwall_origin_{x,y,z}.

Drop-test keys: e_scale_mat{1,4,5,8,9}, rwall_orientation_{rx,ry,rz}.

One-command end-to-end#

After placing the .rad templates in templates/:

cd bumper_beam   # or: cd drop_test
python generate_dataset.py && \
    python run_simulations.py && \
    python ../common/rename_d3plot.py --dataset-dir ./dataset \
        --base-name Bumper_Beam_AP_meshed && \
    python restructure_global_features.py

Handing off to PhysicsNeMo-Curator#

After step 3, the dataset layout matches what PhysicsNeMo-Curator expects:

dataset/
├── run1/
│   ├── d3plot
│   ├── d3plot01
│   └── ...
├── run2/
│   └── ...
└── ...

Point the curator’s etl.source.input_dir at this dataset/ folder and follow the VTP/Zarr export instructions in `examples/structural_mechanics/crash/README.md <../crash/README.md#data-preprocessing>`__. Once curated, the global_features.json produced in step 4 slots directly into the training experiment configs via training.global_features_filepath.

Troubleshooting#

  • ``starter.log`` reports “BAD CARD” or zero E on a ``/MAT`` block. OpenRadioss parses materials with fixed-width 20-char columns. If you add new materials with very large or very small E scales, inspect the rewritten _0000.rad to confirm the scaled value still fits the field.

  • ``anim_to_vtk`` or ``vortex_radioss`` cannot find shared libraries. The runner sets LD_LIBRARY_PATH from OPENRADIOSS_ROOT/extlib/{hm_reader,h3d}/lib/linux64. If your OpenRadioss layout differs, edit common/radioss_runner.build_radioss_env.

  • ``run_simulations.py`` reports “No animation files (A001) found”. The Engine step did not finish a first animation checkpoint — check engine.log. Common cause: timestep collapse after an over-aggressive E reduction on the drop-test flow.

  • ``restructure_global_features.py`` fails with ``Duplicate run_id``. Indicates summary.json was appended to twice. Regenerate from a clean dataset/ folder.

Assumptions & OpenRadioss version notes#

  • Assumed Radioss deck format: block-style .rad files with whitespace- or comma-separated fixed-width fields, matching the Altair docs shipped with OpenRadioss 2024.x. No /INCLUDE or /SUBMODEL support.

  • Material E detection in the drop-test flow relies on a preceding comment line that mentions E and nu/Nu. Remove this comment in the starter and the scaling silently becomes a no-op — intended, since the deck would then have an unknown field layout.

  • The bumper flow scales /PROP/SHELL thickness at line 4, column 3. Decks that place thickness elsewhere (e.g. /PROP/TYPE1) need the mutation logic adapted in bumper_beam/generate_dataset.py.