Training and trust / Trust

Limits and trust

A surrogate is only as good as its data: know the domain, quantify error, verify the winner.

• Extrapolation: Outside the training data, error grows without warning: report the training range with every prediction.
• Data cost: Generating hundreds to thousands of high-fidelity simulations usually dominates the budget.
• Curse of dimensionality: Each additional design parameter multiplies the samples needed to cover the space.
• Interpretability: A neural surrogate is a black box; physics constraints and uncertainty estimates make it auditable.

Related

• Validation: Compare surrogate and solver on designs the network never saw, inside and outside the training range.
• Design use: Millisecond answers turn design studies, optimisation and real-time tools into routine work.
• Technique library: Thirty techniques that make surrogates work in practice, from basics to the frontier.