Case StudyActive

Surrogate Modelling for Structural Dynamics

Exploring how machine learning models can approximate expensive structural dynamics simulations while preserving useful physical behaviour.

Surrogate ModelsStructural DynamicsScientific MLUncertainty

Problem

High-fidelity simulations in structural dynamics can be computationally expensive. This limits rapid design iteration, uncertainty analysis, optimisation, and real-time decision-making.

Surrogate models offer a way to approximate simulation outputs faster, but their usefulness depends on more than low error on a test set. In engineering contexts, the model must remain reliable under changing conditions and physically meaningful inputs.

Core questions

Can a learned model approximate dynamic structural response efficiently?
What types of error matter most in downstream engineering use?
How does the model behave outside the training distribution?
What is the trade-off between speed, accuracy, and reliability?

Approach

This project investigates data-driven surrogate modelling methods for structural response prediction. The work involves simulation data, model training, validation, and analysis of failure modes.

Neural networks
Gaussian processes
Reduced-order modelling
Physics-informed learning
Sequence models
Uncertainty-aware models

Output

Baseline model
Simulation dataset
Training pipeline
Evaluation metrics
Plots comparing predicted and simulated response
Failure analysis notes
Technical write-up

Commercial angle

Reliable surrogate models could support faster design iteration, optimisation, digital twins, structural health monitoring, uncertainty analysis, and engineering decision-support tools.

Next steps

Establish baseline benchmark
Define evaluation metrics beyond simple error
Test robustness under distribution shift
Explore uncertainty estimation

Proof artifacts

Baseline benchmark spec

Next to publish

A compact benchmark definition covering input variables, response targets, train/test splits, and the first error metrics beyond average prediction error.

Prediction vs simulation plot

Planned

A visual comparison of simulated and predicted dynamic response once the first reproducible baseline is stable enough to show.

Failure-mode note

Drafting

A short technical note tracking where the baseline breaks under extrapolation, shifted load cases, and physically implausible inputs.

This is an in-progress research project. Status, methods, and results are exploratory and will be updated as the work develops.