Surrogate-Driven Optimization
A surrogate is a fast learned stand-in for a slow simulation, making optimization and uncertainty analysis affordable at scale.
The bottleneck a surrogate removes
High-fidelity simulations are accurate but slow, so an optimizer that calls them thousands of times is impractical. A surrogate is a model, often a neural network or Gaussian process, trained on simulation results to predict the same output in microseconds. Optimization then runs against the surrogate, with the true simulation used sparingly to check.
Building a good surrogate
- Sample the input space with a design of experiments.
- Run high-fidelity simulations at those samples to build a training set.
- Fit the surrogate and measure its error on held-out points.
- Add samples where the surrogate is uncertain, then refit.
Optimization loop
With a trustworthy surrogate, an optimizer sweeps the input space cheaply. Promising candidates are re-evaluated at full fidelity; if they disagree with the surrogate, those points feed back into training. This active loop concentrates expensive computation where it changes the answer.
Uncertainty comes free
Because a surrogate is cheap to call, it makes uncertainty propagation and global sensitivity analysis affordable, so an optimized design also comes with its error bars.
For Kronos
Surrogates make searching the 25,000+ configuration space and closing the design point tractable for both the Hyperion breeder and the burner. They also run inside real-time control, where full physics is far too slow.
Honesty about limits
A surrogate is only valid where it was trained; extrapolation beyond the training region is flagged and re-checked, never trusted blindly. Its validated range travels with it.