Surrogates and Extrapolation Risk

Fast surrogate models accelerate design but carry a specific risk: confident predictions outside their trained range.

Speed with a caveat

Design and control at Kronos use surrogate models — fast approximations trained on high-fidelity simulations — to explore the design space and to control plasma in real time. Surrogates are powerful because they are orders of magnitude faster than first-principles solvers, but they share a well-known hazard: they can be confidently wrong outside the range of data they were trained on.

Managing extrapolation

• Track the applicability domain: know where the surrogate was trained and flag predictions outside it.
• Attach uncertainty to every surrogate output, so extrapolation shows up as widening error bars.
• Fall back to first-principles solvers for out-of-range or safety-critical decisions.
• Re-validate surrogates as new high-fidelity or, eventually, hardware data arrive.

Why it belongs in the readiness story

Several headline design points involve regimes with limited direct data — D-3He burner performance and net tritium breeding among them. A surrogate that extrapolates into such a regime is making a claim the data do not support, which is exactly the kind of hidden overreach the honest-gates discipline exists to prevent. Kronos treats out-of-range surrogate predictions as hypotheses to be tested, not results, and validates against benchmarks before any number becomes a public claim.

This is the machine-learning face of the broader rule: extrapolation is a gate, not a guarantee.