Design of Experiments
Choosing which points to simulate or test, deliberately rather than randomly, extracts the most knowledge from a limited budget of runs.
The core idea
When each simulation or test is expensive, you cannot run them all. Design of experiments, or DoE, chooses which points to run so that the results reveal the most about the whole space. A good design learns more from fewer runs than a haphazard one.
Common designs
- Factorial designs vary factors systematically to expose main effects and interactions.
- Space-filling designs, like Latin hypercubes, spread points evenly across the input space.
- Sequential designs add points where the current results are least certain.
- Optimal designs place points to minimize a chosen measure of uncertainty.
Space-filling for surrogates
When the goal is to train a surrogate, the design should cover the input space so the surrogate is accurate everywhere it will be used. Space-filling designs do this, and sequential refinement then adds points where the surrogate remains uncertain.
Sequential and adaptive
The most efficient designs are adaptive: run an initial set, learn from it, and choose the next points in light of what was learned. This connects DoE to Bayesian optimization and ML-accelerated experiments, which are DoE turned into a live loop.
For Kronos
DoE organizes the sweep of the configuration space and the plan of irradiation and plasma experiments, ensuring the limited budget of expensive runs maps the space rather than clustering by accident.
Sensitivity link
The same designs feed global sensitivity analysis, so a well-designed sweep answers both what is optimal and what matters.