2. Crash deformation fundamentals
Before any neural network can predict how a vehicle body will fold, stretch, or tear during a collision, we need a precise vocabulary for describing what actually happens to sheet metal and extruded profiles under impact. This section establishes that vocabulary: the dominant deformation modes observed in real crash events, the quantitative metrics engineers use to judge crashworthiness, and why these quantities matter for occupant safety and for the data-driven models developed later in this course.
Deformation Modes in Thin-Walled Structures
Most energy-absorbing vehicle structures — front rails, bumper beams, B-pillars, rocker panels — are thin-walled closed or open sections. When loaded axially or laterally, they deform through a small number of recurring mechanisms. Recognizing these modes is essential because each produces a distinct force–displacement signature, and a prediction model must learn to distinguish them.
- Progressive folding (axial crushing). The classic concertina or diamond mode, where a thin-walled tube forms a sequence of plastic hinges along its length. Force oscillates around a mean crushing force with periodic peaks at each fold. This is the most desirable mode for front rails because it absorbs energy steadily and predictably.
- Global bending. The member buckles as a beam rather than folding locally, producing a single large plastic hinge. Energy absorption is lower and the force–displacement curve is smoother but less efficient. Common when the section is slender or the load is applied off-axis.
- Mixed mode. Local folding and global bending interact, often triggered by geometric imperfections or asymmetric loading. This is the hardest mode to predict and the most sensitive to boundary conditions.
- Fracture and tearing. When local strain exceeds the material's forming limit, cracks initiate at corners or weld lines. Fracture truncates the crushing process and can cause a sharp force drop, reducing total energy absorption.
- Inversion and splitting. The tube wall rolls outward or splits along a seam. These modes appear in specific geometries and are usually avoided by design.
The transition between modes is governed by the ratio of wall thickness to section width, the yield strength and hardening behavior of the material, and the loading angle. A front rail with a thickness-to-width ratio around 0.05 and a high-strength steel grade will typically fold progressively; the same geometry in a softer steel may bend globally.
Crashworthiness Metrics
Deformation modes are qualitative; crashworthiness metrics make them quantitative. The following quantities are extracted from a force–displacement curve obtained from a drop tower test, a sled test, or a validated finite element simulation.
| Metric | Definition | Typical Target (Front Rail Example) |
|---|---|---|
| Peak force (Fmax) | Maximum force during the event | < 120 kN to protect the occupant compartment |
| Mean crushing force (Fmean) | Total absorbed energy divided by crush distance | 60–80 kN for a 300 mm stroke |
| Energy absorption (Eabs) | Area under the force–displacement curve | > 18 kJ for a full front rail |
| Crush force efficiency (CFE) | Fmean / Fmax | > 0.65 indicates stable folding |
| Specific energy absorption (SEA) | Eabs / mass of the member | 15–25 kJ/kg for steel, higher for aluminum |
| Intrusion | Maximum displacement into the occupant cell | < 150 mm at the pedal box |
These metrics are not independent. A design that increases Fmean by thickening the wall may also raise Fmax, reducing CFE and increasing the risk of occupant injury from deceleration spikes. The engineering task is to find a geometry and material combination that keeps Fmax below the injury threshold while maximizing Eabs within the available crush stroke.
Engineering Significance in Vehicle Safety
Crash deformation is not an abstract academic exercise. It directly determines what happens to the occupant. In a frontal impact, the front rail must absorb enough kinetic energy so that the remaining energy reaching the occupant compartment is small enough for the restraint system — seat belts, airbags, and the seat structure — to manage. If the rail fractures prematurely, the force drops, energy absorption stops, and the occupant experiences a second impact against the interior. If the rail is too stiff, the deceleration pulse is too high and the occupant is injured by the belt itself.
This trade-off is why crashworthiness metrics are used as design targets and as validation criteria for any prediction model. A deep learning model that predicts deformation must ultimately output quantities that map to these metrics. Predicting a displacement field is useful only if that field can be integrated to compute Fmean, Eabs, and intrusion. The model's loss function, therefore, often combines a reconstruction error on the deformation field with a physics-based penalty on the derived force–displacement curve.
From Modes to Predictions: A Worked Example
Consider a simplified front rail modeled as a square tube with a 60 mm × 60 mm cross-section, 1.8 mm wall thickness, and a length of 400 mm. The material is a dual-phase steel with a yield strength of 600 MPa and a hardening exponent of 0.15. In a drop tower test with a 500 kg mass impacting at 15 m/s, the measured force–displacement curve shows an initial peak of 95 kN, a mean force of 62 kN over a 280 mm stroke, and a total energy absorption of 17.4 kJ. The CFE is 0.65, indicating stable progressive folding.
If the wall thickness is reduced to 1.4 mm, the same tube may switch to global bending. The peak force drops to 70 kN, but the mean force falls to 38 kN and the energy absorption to 10.6 kJ. The CFE remains around 0.54, but the absolute energy absorption is insufficient for the target. This example illustrates why mode identification is not just descriptive — it is predictive of whether a design will meet its crashworthiness targets.
For a neural network, the input might be a vector of geometric parameters (width, thickness, length, corner radius), material parameters (yield strength, hardening exponent, density), and loading conditions (impact mass, velocity, angle). The output could be a low-dimensional representation of the deformation field, from which the force–displacement curve is reconstructed. The training data would consist of hundreds of such simulations or tests, each labeled with the deformation mode and the crashworthiness metrics. The model learns the mapping from design parameters to deformation behavior, and the metrics provide a physically meaningful evaluation criterion beyond pixel-wise accuracy.
With this foundation — the modes, the metrics, and their engineering meaning — the next step is to examine how these quantities are measured and simulated in practice, and how they become the labels and features for deep learning models.