The Outlier in the Crash

In a large crash study — the kind automotive manufacturers run when validating a new platform — an engineer might review hundreds, sometimes thousands, of finite-element simulations. Each simulation represents a slightly different variant of the crash event: a different material parameter, a slightly different impact angle, a changed component geometry. The goal is to understand the design's robustness across the space of possible real-world conditions.

The problem is not volume. Simulation infrastructure has scaled. The problem is that no one can look at a thousand simulations carefully. Attention distributes. The dangerous outlier — the simulation where the structural response is anomalous, where a beam buckles unexpectedly, where energy is not absorbed as predicted — is the one that slips through.

This is the problem I set out to address. The result became a paper in the International Journal of Crashworthiness, 2022.

What Makes a Crash Simulation an Outlier

The naive answer is: a simulation is an outlier if its deformation is significantly different from the others. But "significantly different" in three-dimensional, time-varying deformation data requires careful definition.

A crash simulation produces, at each time step, a three-dimensional point cloud: the deformed positions of every node in the finite-element mesh. The typical simulation has tens of thousands of nodes and hundreds of time steps. Direct comparison is intractable at scale, and simple summary statistics — peak displacement, absorbed energy — miss the where and how of the deformation.

What we need is a representation that is compact enough to reason about at scale, but expressive enough to capture the geometric structure of the deformation behaviour.

The Method

The approach has two stages.

In the first stage, a geometric deep learning model — specifically, a variant of PointNet++ — learns to encode each time-step geometry into a low-dimensional embedding. The model is trained on the full study corpus in an unsupervised manner, so the representation is shaped by the structure of the data, not by a label.

In the second stage, the sequence of per-timestep embeddings for each simulation is aggregated into a single trajectory representation. Anomaly scores are then computed over this trajectory space using reconstruction-based methods: a simulation is anomalous to the degree that its trajectory cannot be faithfully reconstructed by the trained model.

The Dangerous Case Nobody Noticed

In validation studies, the method consistently surfaces simulations that, on manual review, exhibit genuine structural anomalies: unexpected buckling sequences, atypical load paths, energy absorption distributions that deviate from the study mean in ways that matter for safety assessment.

There is something clarifying about this — and something sobering. The anomaly the model flags first is not usually the largest deviation in energy metrics. It is often a simulation where the spatial pattern of deformation is unusual, even if aggregate quantities appear normal. This is exactly the case human review is worst at detecting: the deviation is not in the number, but in the geometry.

The paper formalises the method, validates it on industrial crash study data in collaboration with Dr. David Kracker, and demonstrates detection performance across multiple study configurations.

What I Learned

Anomaly detection on physical simulation data is different from anomaly detection on tabular or image data in a way I underestimated at the start. The structure is three-dimensional, time-varying, and governed by physics that the model has no explicit access to. The model must infer, from deformation geometry alone, what is normal — and the definition of normal varies across study configurations.

The thing that grounds the work, and that I return to when the methodology feels abstract, is the physical reality at the other end: a crash test, a vehicle, an occupant. The dangerous case nobody noticed is not a data point. It is a design failure waiting to find its way into the real world.

That is the sufficient reason to keep improving the method.