Transport-surrogate validation on two scenarios
The final transport surrogate now has local derivative and conservation checks on ITER hybrid as well as RAPTOR L-mode. Both use the guarded ensemble, with Bohm/gyro-Bohm transport supplying coefficients wherever the guard rejects the surrogate.
Derivatives and conservation
The derivative checks compare automatic differentiation with finite differences. They perturb the electron-temperature profile, require the guard to select the same faces on each side of the perturbation, and require the Newton solver to reach its fine tolerance. Forward and reverse automatic differentiation must also agree. These checks cover both the initial transport coefficients and the core temperature after coupled steps.
The loop checks require finite temperatures, density and poloidal flux. They compare the selected coefficients directly with the ensemble or fallback, then rebuild discrete energy and particle balance from the solver coefficients. The existing assertion tolerances are retained: 15 gradient, JIT, fallback and discrete-balance checks pass across RAPTOR and ITER hybrid in each modeC-146.
ITER hybrid uses an explicit Newton solver override for this validation. The existing linear-solver timing configuration remains a separate experiment. The original surrogate and final model are both checked on RAPTOR; only the final model is checked on ITER hybrid.
What remains open
These are short local checks using ITER ensembles. They do not measure agreement with a fresh TGLF trajectory, establish a final-model speed-up or certify an error bound. The default Newton stopping tolerance still permits derivative and balance discrepancies (K-026); tighter convergence is explicit in these tests.
The new execution reports and tested source snapshots are stored separately and bound to a validation report by hashes. Reports for unchanged test files are retained when generating site statistics; the aggregate count is not a fresh full-suite rerun.
Matched final-model timing still awaits resolution of the compute cancellation. RF and pedestal training still requires the reference-target decision in OQ-21.
Technical details → Extended loop validation.
Technical details → Surrogates › extended loop validation