Limits of ML in Fusion, and Honest Practice
A candid summary of what machine learning can and cannot do for fusion, and the standards that keep it useful.
What ML does well
Machine learning excels at fast interpolation within regimes where data are dense: emulating expensive simulations, estimating state from partial diagnostics, flagging anomalies, and learning control policies against a good simulator. In these roles it complements physics, doing quickly what physics does slowly.
What ML cannot do
- Discover physics absent from its training data
- Extrapolate reliably to genuinely new regimes or machines
- Replace validation against physics and experiment
- Turn a small, biased dataset into confident general knowledge
The extrapolation problem
Fusion's most important questions concern machines and regimes that do not yet exist, exactly where data-driven models are weakest. A model can look confident far outside its training set while being badly wrong. Recognizing and flagging this, through uncertainty quantification and out-of-distribution detection, is the difference between honest and misleading use.
The standards that keep ML honest
Split data by shot and in time; handle imbalance; report calibrated uncertainty and the data range behind every prediction; validate against physics and held-out experiment; and make results reproducible. These are not optional refinements; they are what make a data-driven result an engineering result.
In the Kronos context
The Kronos breeder and burner are design and simulation efforts. Machine learning here accelerates modeling and informs control and design studies; it does not stand in for the physics case, and no hardware net-gain claim is made before first-of-a-kind first tritium. Used within its limits and with its uncertainty stated, ML is a genuine multiplier for fusion modeling. Used beyond them, it produces confident numbers that mean nothing. The discipline is the point.