Determinism vs Learning
Deterministic code does exactly what it was told; learning systems infer what to do from data. Fusion plants need both, kept in their lanes.
Two kinds of software
A deterministic program produces the same output for the same input every time, following rules written by a person. A learning system derives its behavior from data during a training phase, and its rules are implicit in learned parameters. The distinction matters most where consequences are large.
Where determinism belongs
Safety interlocks, machine-protection logic, and control loops that guard expensive hardware should be deterministic and auditable. You must be able to read the logic, prove its behavior, and reproduce it exactly. A magnet quench response cannot depend on a model that might generalize unexpectedly.
Where learning belongs
Learning excels at pattern recognition, estimation of hard-to-measure quantities, and searching large design spaces. These are advisory or offline roles where a human or a deterministic guard reviews the output before it acts.
Keeping the boundary clean
- Learning proposes; deterministic logic disposes.
- Learned outputs are bounded and checked before they reach actuators.
- Every learned component is versioned and reproducible like any other software.
Reproducibility of learning
A learned model can be made deterministic in use: once trained, fixed weights and fixed inputs give fixed outputs. Recording the seed, data, and code that produced those weights restores the reproducibility that engineering demands, so learning need not mean unpredictability.