ML-Accelerated Experiments
Machine learning shortens the loop between running an experiment, understanding it, and choosing the next one.
The experimental loop
Every plasma discharge or irradiation test produces data that should inform the next one. Traditionally that analysis is slow, so many runs are planned in advance and analyzed later. Machine learning speeds the interpretation, letting the next experiment be chosen in light of what the last one just showed.
Where ML helps
- Fast reconstruction of plasma state from raw diagnostics.
- Automatic classification of events and regimes.
- Prediction of an experiment's outcome before it is run.
- Selection of the next most informative experiment.
Active experimentation
Combining prediction with selection turns a fixed test plan into an adaptive one. The system proposes the experiment expected to reduce uncertainty the most, related to Bayesian optimization and to design of experiments. Fewer runs then yield more knowledge.
Training against the twin
Models can be pre-trained against the digital twin before any hardware exists, so that when the Hyperion breeder and the burner begin operating around and after first tritium, the analysis tools are already mature rather than being built from scratch.
Honest validation
An ML prediction guiding an experiment must be validated against real outcomes, and its accuracy tracked openly. A model that guides badly is worse than none, so its performance is monitored continuously through drift detection.
Feedback
Every completed experiment updates the models, tightening the loop and contributing to the intelligence flywheel.