1. Crash deformation fields
A crash deformation field is the spatial description of how a vehicle body or structural component has permanently changed shape after an impact event. It is not a single number or a scalar severity score; it is a field — a function defined over the material domain that assigns, at every point, a measure of how that point moved, stretched, or was displaced relative to its undeformed reference configuration. In geometric deep learning for crash analysis, this field is the central learning target: the quantity we wish to predict, reconstruct, or reason about from partial observations.
Physically, the deformation field arises from the dissipation of kinetic energy through plastic flow, buckling, folding, and fracture. During a frontal offset impact, for example, the front longitudinal rails fold into accordion patterns while the passenger cell may remain largely elastic. The resulting field is therefore highly non-uniform: large displacements and rotations concentrate in energy-absorbing members, while rigid or protected regions show near-zero deformation. Understanding this spatial heterogeneity is essential, because a model that treats the whole body as a single rigid transformation will fail precisely where the physics matters most.
Reference and Current Configurations
Every deformation field is defined relative to a reference configuration, typically the as-designed CAD geometry or a pre-crash scan. Let X denote a material point in the reference configuration and x its position after the crash. The displacement field is then u(X) = x − X. This convention — displacement measured from reference to current position — must be fixed at the outset and held consistently throughout any pipeline, because sign errors and frame mismatches are among the most common causes of silent training failures.
Displacement alone does not capture stretching. The deformation gradient F = ∂x/∂X = I + ∂u/∂X encodes local stretch and rotation. From it we derive the Green–Lagrange strain tensor E = ½(FTF − I), which is invariant to rigid-body rotation and therefore preferable to small-strain approximations when rotations are large — as they invariably are in crash events. For thin sheet-metal panels, engineers often work with principal strains or the equivalent plastic strain, since these correlate directly with material failure and are the quantities used in forming-limit and crashworthiness criteria.
Measurement Modalities
Deformation fields reach the learning pipeline through several modalities, each with distinct resolution, noise, and coverage characteristics.
- Full-scale crash tests with optical tracking: Stereo or multi-camera photogrammetry tracks painted targets on the body, yielding sparse 3D displacement histories at discrete points over time. Dense digital image correlation (DIC) can recover surface fields but is sensitive to lighting, occlusion, and high-speed motion blur.
- Post-crash 3D scanning: Structured-light or laser scanners produce dense point clouds of the deformed body. Registration against the reference CAD or pre-crash scan yields displacement and strain fields, but only at the final state — temporal history is lost.
- Component-level tests: Quasi-static or dynamic axial crushing of rails and bumper beams provides controlled, well-instrumented strain data, often used to calibrate material models and to generate training data at higher fidelity than full-vehicle tests.
- Finite element simulation: Explicit solvers such as LS-DYNA or Radioss produce dense, noise-free fields at every time step. These are the dominant source of training data, but they carry modeling bias and must be validated against physical tests.
Representations: Displacement, Strain, and Point Clouds
The same physical deformation can be represented in several ways, and the choice strongly constrains which geometric deep learning architecture is appropriate.
| Representation | What it stores | Strengths | Limitations |
|---|---|---|---|
| Displacement field | Per-vertex or per-point vector u | Directly interpretable; easy to visualize; natural output of registration | Not rotation-invariant; large rigid motions dominate the signal |
| Strain field | Tensor or scalar (e.g., equivalent plastic strain) per element | Rotation-invariant; physically meaningful for failure | Requires differentiation; noisy on sparse data; tensor-valued |
| Point cloud | Unstructured set of 3D coordinates | Native scanner output; permutation-invariant architectures apply | No connectivity; correspondence to reference must be established |
In practice these representations are coupled. A point cloud is registered to the reference mesh to obtain correspondences, displacements are computed per vertex, and strains are derived by differentiating the displacement field over the mesh connectivity. Each step introduces error, so a well-designed pipeline documents where each representation is produced and which one the network actually consumes.
From Measurement to a Learnable Field
Consider a concrete example. A post-crash scan of a B-pillar yields roughly 400,000 points. The reference CAD mesh has 60,000 vertices. Registration via coherent point drift or a learned feature-matching network establishes correspondences, and displacement is computed at each matched vertex. Because the scan is noisy, a smoothing step — for instance, projecting displacements onto a low-order spectral basis of the mesh Laplacian — suppresses high-frequency noise while preserving the large folds that carry the physics. The resulting per-vertex displacement field, together with the mesh connectivity, is the input to a graph neural network or a mesh convolutional operator.
The figure below summarizes how the physical event, the measurement modalities, and the three representations connect, and where the learning target sits.
Why Representation Choice Matters for Learning
The representation determines the inductive bias the network needs. A displacement field is a vector-valued function on the mesh; a graph neural network with edge features can propagate it across connectivity, but it must learn to separate rigid motion from genuine deformation. A strain field is scalar or tensor-valued and rotation-invariant, which reduces the burden on the model but requires a differentiable mapping from displacement — often implemented as a fixed finite-element or finite-difference operator applied inside the network. A point cloud discards connectivity, so permutation-invariant architectures such as PointNet-style encoders or transformer-based set models are natural, but correspondence to the reference must be recovered separately if per-vertex displacement is required.
A useful diagnostic when building such a pipeline: verify that a rigid translation applied to the entire deformed body leaves the strain field unchanged while shifting the displacement field uniformly. If your implementation violates this, the reference frame or the differentiation operator is wrong. Similarly, check that a pure rotation produces zero Green–Lagrange strain; a non-zero result indicates a linearized strain approximation being used where rotations are large.
With the physical meaning, the measurement modalities, and the three representations established, the stage is set for the geometric learning machinery itself: how to define convolution, attention, and message passing on meshes and point clouds so that the deformation field can be learned efficiently and generalizes across vehicle architectures and impact configurations.