~/career/phd

PhD

Location 52.4227° N · 10.7865° E 52°25′22″ N 10°47′11″ E
scroll

Dissertation

My PhD thesis focuses on knowledge management in CAE, specifically in the field of simulation-based crashworthiness design. As virtual testing continues to expand, simulations generate increasingly large and complex datasets across a growing number of load cases, vehicle variants and level of detail, while the crash behaviour of the overall vehicle structure remains highly nonlinear — creating both methodological challenges, such as the curse of dimensionality, and organizational challenges related to the tool landscapes to handle the amount of generated data. My goal is to address these challenges by applying data analytics methods to structure, analyse and connect simulation data, making engineering knowledge easier to extract, reuse and leverage in the design process.

Complication

What happens in a frontal crash is the result of many coupled nonlinearities. The FE model starts moving – scroll through the simulation.

ComplicationLoad case

Frontal crash against a rigid wall

A standard load case in crashworthiness design – and a good example of how much happens within a few milliseconds.

ComplicationImpact

The structure deforms

The crash boxes fold, one longitudinal rail buckles, the hood bulges and the rear lifts. Forces flow along three load paths; the colour contour shows the plastic strain (schematic).

  • Load path 1 · crash box → longitudinal rail → sill
  • Load path 2 · upper rail (shotgun) → A-pillar → roof
  • Load path 3 · subframe → tunnel
ComplicationNonlinearity 1/5

Elastic-plastic material behaviour

The material is loaded beyond its yield point: on the flow curve, the state moves from the elastic into the plastic range – the deformation remains.

elastic plastic σ ε yield point
ComplicationNonlinearity 2/5

Buckling of profiles

Under axial compression, one longitudinal rail buckles to one side: a plastic hinge forms – a single kink instead of even folding. Small imperfections decide whether and where this happens.

F axial compression → buckling (plastic hinge)
ComplicationNonlinearity 3/5

Material failure

At the joint between longitudinal rail and sill, elements exceed their failure limit and are deleted – a crack runs around the profile and the load path is interrupted.

element failure → elements deleted, crack
ComplicationNonlinearity 4/5

Component collisions

Late in the crash, the engine tilts backwards and hits the master brake cylinder. The brake booster tilts and strikes the firewall with its edge – a local intrusion.

booster firewall brake booster ← engine
ComplicationNonlinearity 5/5

Modelling errors

Contacts can snag numerically: nodes penetrate surfaces and components drag each other along – an effect that does not exist in reality.

! penetration → contact snags, parts drag along
ComplicationOverview

Everything happens at once

This makes predictions very difficult: there are countless effects and interactions in the structure – and they influence each other.

  • 1Elastic-plastic material behaviour
  • 2Buckling of profiles
  • 3Material failure
  • 4Component collisions
  • 5Modelling errors
SolutionData lake

The data is already there

We have vast amounts of simulation data – mostly stored in data lakes – and a large toolbox of machine learning and data analytics methods.

Unlocking the treasure

Now it's time to unlock this treasure with new methods of data analytics and AI – and to establish data-driven development processes.

We don't lack data – and the expert knowledge to feed AI assistants is buried in it.

Anomaly detection

Learn what is normal – flag what is not

Each point is one simulation run (schematic). Clusters are the typical deformation modes; runs that deviate are flagged automatically – this is exactly where the knowledge hides.

  • normal: known deformation modes
  • slightly unusual
  • anomaly: significant increase in intrusion, unexpected buckling, structural integrity

Schematic example: one modification, one condition, one observed effect. Switch the variant and watch the structure in the scene.

X
Geometry of the longitudinal rail changed
Y
Load case: frontal crash, model variant […]
Z
Changed firewall intrusion
Intrusion

A Knowledge Graph of engineering experience

The key step is what happens next: these individual observations are not stored as isolated records. They are connected in a Knowledge Graph. Over time, the graph builds relationships between components, modifications, physical effects, load cases, model variants and previous engineering decisions.

In this way, thousands of individual simulation runs gradually form a connected representation of engineering experience. This creates a particularly powerful foundation for AI.

Asking the assistant

Imagine an engineer designing a new structure who observes an unwanted deformation pattern. Instead of manually searching old projects, the engineer asks an AI assistant.

engineer> "Have we seen this behaviour before, and which design changes improved it?"

assistant> navigating knowledge graph …

  1. current problem
  2. similar physical effects
  3. previous components & projects
  4. successful modifications
  5. supporting simulations

assistant> [illustrative answer – e.g. relevant earlier variants and their supporting simulations]

Talks & Publications

  • Ein intelligenter Assistent für die Crashauslegung – Perspektiven und Voraussetzungen

    [An intelligent assistant for crashworthiness design – perspectives and prerequisites] · 21st VDI Conference „SIMVEC – Auslegung und Absicherung von Fahrzeugsystemen“, Baden-Baden, Germany

    doi.org/10.51202/9783181024454-415
  • An Intelligent Assistant for Automotive Design for Crashworthiness – Development and Application Insights

    Artificial Intelligence and Machine Learning in CAE Applications Conference, Munich, Germany

    nafems.org – AI/ML in Engineering Simulation
cd ../master-thesis