# Build AI Models for Crash Analysis with NVIDIA PhysicsNeMo

## Executive summary

This video details the use of AI surrogate models, specifically within the NVIDIA PhysicsNeMo framework, to accelerate structural and automotive crash analysis. By training models on high-fidelity simulation data, engineers can quickly predict how new designs will deform, significantly speeding up the design iteration process. The workflow covers data preparation (using DOE), model architecture (including Transolver and GALE Attention), and comparing various prediction strategies (one-shot, time-conditioned, and auto-regressive).

## Key takeaways

- AI Surrogate Modeling for Crash Analysis: AI models learn from complex, time-consuming crash simulations (which involve large deformation, bending, and contact) to rapidly predict structural deformation for new designs, allowing engineers to explore more ideas faster. (00:00)
- Data Preparation and Validation: Datasets are created using Design of Experiment (DOE). Validation sets are used to measure generalization and set acceptable error limits, while an optional guardrail flags unusual inputs or geometries for manual review. (00:57)
- Advanced Prediction Methods: Three prediction methods are compared: One-shot (predicts future in one call), Time-conditioned (predicts each time step independently), and Autoregressive (advances step-by-step, risking error accumulation). (02:15)

## Technical details

- Data Workflow and Formats: Simulation results (position, displacement, stress, plastic strain) are extracted from native formats into structured inputs like VTK, ZAR, or PhysicsNeMo mesh for AI training. (01:36)
- PhysicsNeMo Workflow: The PhysicsNeMo crash recipe provides a modular workflow managed by Hydra configuration, supporting multiple architectures and time integration methods. (02:53)
- Model Architecture Concepts: The Transolver theory is introduced, explaining how prepared data becomes model input. GALE Attention is used to summarize information from the mesh and connect structural state to overall crash conditions. (04:45)
- Educational Notebook Structure: The workflow is broken down into three notebooks: Notebook 0 (Data Preparation, starting with native VTP files), Notebook 1 (Model Architecture/Theory), and Notebook 2 (Training, Inference, and Evaluation). (03:50)

## Practical implications

- Accelerates the design cycle for structural components by providing rapid, predictive analysis of crash performance.
- Allows engineers to evaluate a greater number of design iterations compared to relying solely on full-scale physical testing or simulations.
- The modular workflow (PhysicsNeMo) enables comparison and selection among different time integration and prediction strategies.

## Topics

AI Surrogate Modeling, Structural Mechanics, Automotive Crash Analysis, Deep Learning, Physics-Informed AI, PhysicsNeMo crash training recipe, Crash-analysis notebooks, PhysicsNeMo blog post

Source: https://www.youtube.com/watch?v=ZpjB-y1l1kg