Machine learning surrogate models for CFD and FEA: where they pay off
A trained surrogate can answer in milliseconds what a solver answers in hours. When that trade is worth making, how to build one that holds up, and the warning signs that it will not.
Surrogate models replace an expensive simulation with a fast approximation learned from its results. Done well, they turn a design study of a few dozen CFD or FEA runs into an exploration of thousands of options. Done badly, they produce confident answers with no physics behind them.
The idea is not new: response surfaces and kriging have supported design of experiments for decades. What has changed is that modern machine learning can learn whole fields rather than single numbers, and can be combined with physical laws. That widens where surrogate models are useful, but it does not remove the need for engineering judgement.
When surrogate models are worth building
A surrogate costs effort twice: generating training data and validating the result. It pays off when the same kind of question will be asked many times.
- Design optimisation over many parameters, where an optimiser needs thousands of evaluations.
- Uncertainty quantification, where manufacturing and operating variation must be propagated through the model.
- Real-time use: digital twins, control design and operator tools that cannot wait for a solver.
- Product families, where the same geometry is re-run with different dimensions for each customer.
When not to build one
Types of surrogate models
| Method | Predicts | Strength | Watch out for |
|---|---|---|---|
| Polynomial response surface | A few scalar outputs | Transparent and cheap | Misses strong non-linearity |
| Gaussian process (kriging) | Scalars with an uncertainty estimate | Works with little data, gives error bars | Scales poorly beyond a few thousand samples |
| Neural network | Scalars or whole fields | Handles many inputs and non-linear behaviour | Needs more data; no built-in uncertainty |
| Reduced-order model (POD and similar) | Fields from a reduced basis | Keeps the structure of the solution | Struggles with moving features such as shocks |
| Physics-informed network | Fields constrained by governing equations | Uses physics to reduce data needs | Training can be slow and delicate |
Building one that holds up
- Define the design space honestly. Bounds on every parameter, and the operating conditions that will actually be queried.
- Sample it well. Space-filling designs such as Latin hypercube, then adaptive sampling where the surrogate is least sure.
- Keep the simulations consistent. Same mesh strategy, solver settings and convergence criteria for every sample, or the surrogate learns numerical noise.
- Hold data back. Validate on runs the surrogate never saw, including points near the edges of the space.
- Report uncertainty. Prefer methods or ensembles that say how sure they are, and route low-confidence queries back to the solver.
Physics-informed and hybrid approaches
Physics-informed neural networks add the residual of the governing equations to the training loss, so the network is penalised for violating conservation laws as well as for missing the data. The approach was popularised by Raissi, Perdikaris and Karniadakis (2019) and is an active research area. In industrial practice, hybrid models are often more robust: the solver handles what it does well, and machine learning corrects or accelerates a specific part, such as a turbulence closure or a material law.
We describe these options in more depth under physics-informed machine learning and hybrid physics-AI modelling.
Warning signs in a surrogate
- Excellent accuracy on training points and much worse on held-out ones: over-fitting.
- Smooth predictions across a region where the physics changes regime, such as flow separation or buckling.
- Queries far outside the sampled space answered with the same confidence as those inside it.
- Validation only on averaged scalars when the design depends on a local peak.
Getting started
The best first project is a design question your team asks repeatedly and already answers with a trusted, well-validated simulation. Wrap that simulation in a sampling plan, build a modest surrogate, and measure the time saved against a few confirming solver runs. Our work on surrogate modelling and reduced-order models and uncertainty quantification follows exactly that path, from turbomachinery to structural design.
