Guards and data interfaces for RF and pedestal surrogates
The next M3 modelling tasks are RF deposition and the pedestal boundary. They need different outputs from turbulent transport: heating profiles in the plasma volume, and values describing the edge. Their software interfaces are now built.
A complete source profile
The RF guard checks inputs, output bounds and ensemble spread. Its default decision selects a complete source profile. Mixing surrogate values at some radii with fallback values at others can change the integrated heating power, even when each profile separately has the right power. An explicit physics check can reject a candidate whose volume integral is wrong. Selecting the whole profile does not itself enforce power conservation.
A separate vector guard handles boundary responses, with units and physics criteria supplied by the caller. Both guards retain a Tier A fallback and isolate rejected surrogate branches from derivatives. The core loop records source spread and fallback use.
The dataset interface stores case responses or complete radial profiles. Related launches, equilibria or trajectories share a group identifier and stay together in training or holdout sets. Source hashes, units, reference version, failures and the split seed are recorded. Radial profiles keep their supplied toroidal-flux coordinate; there is no silent conversion from poloidal flux.
The synthetic software checks give 25 synthetic contract tests pass in full and Tier-A-only modesC-143. Nonfinite Gaussian RF parameters are also rejected: 12 nonfinite-parameter cases pass in each modeC-144. These tests establish interface behaviour, not prediction accuracy for an RF or pedestal model.
The reference comes next
The existing RF baseline is an analytic Gaussian EC source. The pedestal baseline takes prescribed values. Training against these would demonstrate a pipeline, but would not add predictive physics. Choosing reference solvers or data is the next modelling decision.
After that choice: label grouped cases, inspect a baseline, set accuracy targets, train the ensembles, integrate them into TORAX, and check coupled derivatives, balance and cost. No RF/pedestal model has been trained in this step.
Technical details → RF and pedestal interfaces.
Technical details → Surrogates › rf pedestal