Add OpenRadioss Bumper Beam Community Dataset Section to Crash Example README.md

examples/structural_mechanics/crash/README.md

@@ -637,6 +637,125 @@ Muon: Car-crash test MSE at probe location (Driver, Passenger):
 
 </p>
 
+## Community Dataset: OpenRadioss Bumper Beam
+
+This section describes an open-source dataset generated with
+[OpenRadioss](https://github.com/OpenRadioss/OpenRadioss) and contributed by the community
+to provide a freely accessible alternative for experimenting with this example.
+
+### Dataset Overview
+
+<p align="center">
+<img src="https://github.com/user-attachments/assets/d1f935a5-11ef-4519-baf4-6a1f94c42fcf" width="400">
+</p> 
+
+The dataset is based on the
+[OpenRadioss Bumper Beam example](http://openradioss.atlassian.net/wiki/spaces/OPENRADIOSS/pages/11075585/Bumper+Beam),
+a standard crash benchmark consisting of a bumper beam impacting a rigid cylindrical pole.
+Raw simulation data was generated by iteratively varying two design parameters across 131 runs:
+
+| Parameter | Range | Description |
+|---|---|---|
+| Shell thickness | 1.2 mm – 2.0 mm | Applied uniformly to `PROP/SHELL` IDs 2 and 7 |
+| Pole Y-axis offset | −500 mm – +500 mm | Shifts the rigid pole impact position along the beam |
+
+
+OpenRadioss ANIM output files are first converted to **d3plot** format using
+[Vortex-Radioss](https://github.com/Vortex-CAE/Vortex-Radioss), then curated to **VTP** format
+using [PhysicsNeMo-Curator](https://github.com/NVIDIA/physicsnemo-curator) for training.
+
+The dataset (131 runs, including both raw d3plot and curated VTP files) is publicly available
+on Hugging Face:
+
+> 🤗 **[AIRBORNEPANDA/BumperBeamCrashExample](https://huggingface.co/datasets/AIRBORNEPANDA/BumperBeamCrashExample)**
+
+### Generating the Dataset Yourself
+
+A full data generation pipeline is provided at
+**[HoussemMouradi/OpenRadioss2PhysicsNeMo](https://github.com/HoussemMouradi/OpenRadioss2PhysicsNeMo)**.
+It automates the complete workflow: modifying OpenRadioss input decks, running the solver,
+converting ANIM → d3plot, and generating `GLOBAL_FEATURES.json`.
+
+Interactive Jupyter notebooks walk through every step and handle all dependency installation
+automatically:
+
+| Platform | Notebook |
+|---|---|
+| Windows (local Jupyter) | [OpenRadioss2PhysicsNEMO_windows.ipynb](https://github.com/HoussemMouradi/OpenRadioss2PhysicsNeMo/tree/main/OpenRadioss2PhysicsNeMo_NotebooksOpenRadioss2PhysicsNeMo_windows.ipynb) |
+| Linux (local Jupyter) | [OpenRadioss_PhysicsNEMO_linux.ipynb](https://github.com/HoussemMouradi/OpenRadioss2PhysicsNeMo/tree/main/OpenRadioss2PhysicsNeMo_Notebooks/OpenRadioss2PhysicsNeMo_linux.ipynb) |
+| Google Colab ☁️ | <a href="https://colab.research.google.com/gist/HoussemMouradi/6e88c23e0b272621de0aa46c003c027e/openradioss2physicsnemo_linux.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a> |
+
+After running a notebook, follow the
+[PhysicsNeMo-Curator crash ETL instructions](https://github.com/NVIDIA/physicsnemo-curator/blob/main/examples/structural_mechanics/crash/README.md)
+to curate the d3plot outputs to VTP format for training.
+
+### Training Configuration
+
+The following configuration was used to train on the Bumper Beam dataset. It is based on the
+[`bumper_geotransolver_oneshot.yaml`](https://github.com/NVIDIA/physicsnemo/blob/main/examples/structural_mechanics/crash/conf/bumper_geotransolver_oneshot.yaml)
+experiment preset, using the **VTP reader**, **GeoTransolver (one-shot)** model, and
+**point cloud** datapipe.
+
+Run with:
+
+```bash
+python train.py --config-name=bumper_geotransolver_oneshot 
+```
+
+The key configuration values are:
+
+```yaml
+# Bumper beam crash experiment using GeoTransolver (one-shot).
+# Usage: python train.py --config-name=bumper_geotransolver_oneshot
+
+experiment_name: "Bumper-GeoTransolver"
+
+defaults:
+  - reader: vtp
+  - datapipe: point_cloud
+  - model: geotransolver_one_shot
+  - training: default
+  - inference: default
+  - _self_
+# -- Data ----------------------------------------------------------------------
+training:
+  raw_data_dir: ./CURATED_DATA_VTP/TRAINING_DATA                # path to curated VTP train split
+  raw_data_dir_validation: ./CURATED_DATA_VTP/VALIDATION_DATA    # path to curated VTP validation split
+  global_features_filepath: ./CURATED_DATA_VTP/GLOBAL_FEATURES.json   # path to GLOBAL_FEATURES.json
+  optimizer: muon
+
+# -- Dataset -------------------------------------------------------------------
+num_time_steps: 11
+num_training_samples: 124
+num_validation_samples: 7
+
+inference:
+  raw_data_dir_test: ???       
+
+# -- Datapipe features ---------------------------------------------------------
+datapipe:
+  static_features: [thickness]             # per-node static features for this dataset
+  dynamic_targets:
+    - effective_plastic_strain
+    - stress_vm
+  global_features:
+    - thickness_scale             # shell thickness (mm)
+    - velocity_x                  # initial impact velocity in X (mm/ms)
+    - rwall_origin_y              # rigid pole Y-coordinate (mm)
+  sample_type: all_time_steps
+
+# -- Model ---------------------------------------------------------------------
+model:
+  functional_dim: 4               # input coords (x, y, z) + static_features (1)
+  out_dim: 50                    # (num_time_steps - 1) * 5 = 50 * 5
+  global_dim: 3                   # must match len(datapipe.global_features)
+```
+
+### Results
+<img width="480" height="206" alt="1769108712488" src="https://github.com/user-attachments/assets/53883c93-fe3f-4fb2-8d10-f474693137f3" />
+<img width="480" height="206" alt="1769108706907" src="https://github.com/user-attachments/assets/85a5ed85-f835-45d7-a6b6-86ed875163c6" />
+
+
 ## TODO
 
 - [ ] **Normalize global features**: Global features (e.g., velocity_x, thickness_scale, rwall_origin_y) are currently passed to the model without normalization. Add support for computing and applying per-feature mean/std (or similar) so global inputs are normalized consistently with node features and positions.