Outliers in the Crash

Large-scale crash studies produce thousands of finite-element simulations. Monitoring them by hand is slow and inconsistent; the dangerous case is the one nobody noticed.

I built methods that learn a low-dimensional representation of each simulation and detect anomalous deformation behaviour automatically — geometric deep learning on the mesh, anomaly scoring on the embeddings.

The work became a method, then a paper, then production tooling that surfaces the simulation an engineer needs to look at.