Active Learning for Expensive Simulations
Spending a limited simulation budget on the points that most improve a model or answer a question.
The idea
When each simulation is expensive, choosing which to run matters as much as the model fit to the results. Active learning selects the next inputs to evaluate so that each run adds the most information, rather than sampling blindly or on a fixed grid.
Selection strategies
- Uncertainty sampling: run where the current model is least sure
- Query by committee: run where an ensemble of models disagrees
- Expected model change: run where a point would most update the model
- Task-driven: run where a point most improves a downstream decision
Why it pays off in fusion
Gyrokinetic and neutronics runs, and physical experiments, cost dearly. A space-filling grid wastes runs on regions the model already handles and under-samples the tricky ones, such as transport thresholds. Active learning concentrates effort where it changes the answer.
Batch and cost awareness
Simulations often run in parallel batches, so practical active learning picks diverse batches rather than one point at a time. When different runs cost different amounts, cost-aware variants trade information gained against compute spent, seeking the best information per unit cost.
Cautions
Active learning tied to a flawed uncertainty estimate can chase the wrong points; the model's sense of what it does not know must itself be reliable, which is one reason Gaussian processes pair well with it. And a model built by active learning is still only valid inside the region it explored. Used with these caveats, it stretches a scarce simulation budget substantially.