Ms. Thesis: Graph Neural Networks for Occupant Crash Simulation

Graph Neural Networks for Occupant Crash Simulation:

Train in One Vehicle, Use in Another

Autoliv develops world-leading occupant safety systems and relies extensively on finite element (FE) simulations to evaluate

restraint performance. These simulations are accurate but computationally expensive, which limits the number of load cases,

occupant sizes, and vehicle interiors that can be explored.

Earlier studies have shown that machine learning (ML) can predict occupant kinematics and injury metrics from crash parameters

(Niranjan Poojary & Srinivas, 2021; Wang & Xu, 2022). These ML surrogates map a fixed set of input parameters to outputs and are

therefore tied to the environment they were trained in.

A common way to widen the scope is to use a generic sled model with parameterized vehicle interior (Iraeus & Lindquist, 2016).

This covers more vehicles, but only those that can be described by the chosen parameters, and a surrogate trained on it still

learns a mapping from those parameters.

In parallel, graph neural networks (GNN) have matured. They have been used as autoencoders (Fichera, C. R. 2022) in ML pipeline.

MeshGraphNet (Pfaff et al., 2021) and related graph networks handle interaction and contact between objects (Allen et al., 2023),

and graph and transformer surrogates such as GeoTransolver (Adams et al., 2025) have shown promising results for full-field crash

deformation (Curtosi et al., 2026). A contact-aware GNN offers a way around the parameter limitation: the interior enters as

geometry instead of parameters, so the model learns the local interaction between the occupant and the interior surfaces, which

does not depend on a specific vehicle.

This leads to the key question:

Can a GNN trained in one vehicle environment predict the occupant response in another?

This thesis aims to develop and evaluate such a transferable surrogate using an existing Autoliv dataset of sled simulations with

crash test dummies (ATDs). This means a new interior can be used without retraining, and the method can be tested by training on

one vehicle interior and evaluating on an unseen one.

The thesis is expected to deliver a prototype GNN that predicts ATD kinematics and interior contact in a sled environment,

together with a quantified answer on how well it transfers to a new vehicle interior. The results will show which outputs can be

predicted reliably, and which interior environments can be handled by the method.