Surrogate Models and Machine Learning
Surrogate models approximate expensive physics codes with fast learned functions, enabling real-time control, large scans, and optimization at reduced fidelity.
Trading fidelity for speed
Many fusion codes are too slow for tasks that need thousands or millions of evaluations: real-time control, design optimization, or uncertainty quantification. A surrogate model is a fast approximation, trained on a database of full-code runs, that reproduces the code's input-output behavior at a fraction of the cost.
How surrogates are built
- Generate a training set by running the full code across the input space
- Fit a fast model: regression, Gaussian processes, or neural networks
- Test the surrogate against held-out runs to measure its error
- Deploy it where speed matters, within its validated input range
Where they help
Surrogates for turbulent transport let predictive simulations run quickly; surrogates for equilibrium or heating enable real-time control; surrogates for expensive neutronics or gyrokinetics make broad design scans and optimization feasible. In each case they extend the reach of a slow but accurate code.
The trust boundary
A surrogate is only reliable inside the region of input space it was trained on. Extrapolation beyond that region can give confidently wrong answers, so a well-built surrogate reports when an input falls outside its training domain. Its error against the full code must be quantified and stated, not assumed small.
Machine learning in the pipeline
Beyond surrogates, machine learning is used for disruption prediction, anomaly detection in diagnostics, and pattern discovery in large databases. In all cases the learned model is a tool within a physics-grounded workflow, validated against physics and data, not a replacement for understanding.
Used with discipline, surrogates and learned models accelerate analysis while the underlying physics codes remain the ground truth against which they are checked.