CrashSurrogate-AI

Surrogate Modeling for Bumper-Beam Crash Response using OpenRadioss Data

This project demonstrates a surrogate-modeling workflow for automotive crash simulation data using an OpenRadioss bumper-beam dataset.

The goal is to predict crash-response metrics from simulation input parameters using machine learning and deep learning models. The project focuses on predicting maximum bumper-beam displacement from OpenRadioss simulation results.

---

Project Motivation

Full finite element crash simulations can be computationally expensive and time-consuming. Surrogate models can help approximate simulation responses faster, supporting design-space exploration and early engineering decision-making.

This project was developed as a portfolio-level demonstrator to show skills in:

• crash simulation data handling
• VTP result-file processing
• engineering feature extraction
• surrogate modeling
• machine learning model comparison
• basic deep learning with PyTorch

---

Dataset

The project uses an OpenRadioss bumper-beam simulation dataset hosted on Hugging Face:

Dataset: AIRBORNEPANDA/BumperBeamCrashExample The dataset contains OpenRadioss bumper-beam simulation results in VTP format. The global metadata file contains 131 simulation cases. Seven VTP result files were unavailable, so the final processed dataset contains 124 valid simulation runs.

Input parameters

The available global input parameters are:

• thickness_scale
• velocity_x
• rwall_origin_y

In the processed dataset, velocity_x is constant, so the main learning features are:

• thickness_scale
• rwall_origin_y

Extracted response metrics

From each VTP result file, the following response metrics were extracted at the final time step:

• maximum displacement
• mean displacement
• maximum von Mises stress
• mean von Mises stress
• maximum effective plastic strain
• mean effective plastic strain

The main surrogate-modeling target is:

max_displacement_t100