Surrogate Models
A surrogate model is a fast approximation of a slow simulation, trained to reproduce its outputs across a design space.
Trading fidelity for speed
High-fidelity simulations can take hours or days per run, which is too slow for tasks that need thousands of evaluations, such as optimization or real-time estimation. A surrogate model learns the input–output relationship of the expensive simulator and evaluates it in a fraction of the time.
How surrogates are built
- Sample the expensive model across the input space.
- Fit a fast approximator — a polynomial, a Gaussian process, or a neural network — to those samples.
- Validate the surrogate against held-out simulator runs.
- Deploy it where speed matters, with its error bounds attached.
The domain-of-validity rule
A surrogate is only trustworthy within the region where it was trained. Extrapolation beyond that region can be confidently wrong. Responsible use includes detecting when a query falls outside the training envelope and falling back to the full simulator.
Where they earn their keep
Surrogates enable rapid design-space exploration, sensitivity studies, and real-time estimation of quantities that are otherwise too slow to compute live. In a fusion control setting, a surrogate can predict plasma response fast enough to inform a control loop that the full physics model could never keep up with.
Physics-grounded surrogates
A surrogate trained on a verified first-principles model inherits physical structure, which helps it generalize and keeps its errors interpretable. This coupling of physics and speed is a recurring pattern in the intelligence layer of a fusion plant.