MAST-U remediation starts with matched fluxes and guard history
The owner prioritized remediation of the failed MAST-U coupled transport check. The chosen temperature limit stays in place. Before changing a model, this step checks which predictions are wrong and whether the guard explains the observed fallback faces.
Matched fluxes. A bounded local probe sends the failed run’s final-state inputs to TGLF, using the existing solver settings. Every in-domain face returns finite fluxes. The isolated inner served face contributes most of the temperature error, yet both the surrogate and TGLF predict zero heat flux there at the final state. That result shows why a local final-state prediction cannot explain the whole evolving profile. Elsewhere, flux discrepancies are large: all 17 in-domain final-state probes succeed; on 12 served faces with reference flux above 1 gB, median relative errors are 41.17% Q_e and 74.27% Q_i; an unconstrained v3 control gives 38.05/71.88%C-181.
The existing unconstrained model is a diagnostic control, evaluated on those same states. It also has large errors. This comparison provides no demonstrated shortcut through dropping the owner-required monotonicity constraint.
| diagnostic | result |
|---|---|
| matched flux probes and diagnostic control | all 17 in-domain final-state probes succeed; on 12 served faces with reference flux above 1 gB, median relative errors are 41.17% Q_e and 74.27% Q_i; an unconstrained v3 control gives 38.05/71.88%C-181 |
| training use of the failed-case probes | none |
| replay | all saved temperature, density and psi values are reproduced exactly; four final inner faces exceed the existing ion-diffusivity guard bound; 44.81% of accepted-state face diagnostics reject the surrogateC-182 |
Guard history. The local surrogate replay captures diagnostics at accepted states, after timed stepping. At the final state, four inner faces are inside the input box but have predicted ion diffusivity above the existing guard bound. The served region changes during the run. These diagnostics still do not count intermediate evaluations inside the solver, and the guard settings are unchanged.
Next experiment. A prepared label-refresh pipeline collects surrogate-feedback states on training equilibria, filters them through the existing approved region and keeps the full solver inputs. Both test equilibria remain excluded. A training-case smoke run and focused filtering checks pass; no production label job has been submitted. Its compute request goes to the owner before execution. Retraining and another coupled comparison will be needed to establish whether the proposed coverage remedy works.
Technical details → Verification › mastu remediation
Decisions → OQ-25