M2 · Training data
Done 3 Oct 2026
Goal
Produce the data the surrogates of M3 are trained on: fluxes from a first-principles turbulence model, sampled over the operating space of each machine, stored as immutable, versioned datasets with provenance. The brief requires the sampling to cover spherical-tokamak conditions: the transport network used in M1, QLKNN, was trained at a fixed aspect ratio of 3 and does not.
Acceptance criteria (fixed in the brief before work started)
- The dataset loads.
- Its provenance validates.
- The sampled domain is documented and plotted against the intended operating space.
M2 has no human gate in the brief. The kinetic ranges were approved by the owner before any dataset was generated (OQ-16). All three criteria were met on 3 Oct 2026, and the owner closed M2 (OQ-18).
Tasks
| task | state | log |
|---|---|---|
| turbulence code: TGLF (GACODE, Apache-2.0) built, checked, timed; CGYRO built for spot checks | done | entry 16 |
| geometry ranges measured on MAST-U (FreeGSNKE) and ITER (TORAX) equilibria | done (OQ-15) | entry 17 |
| kinetic ranges, species and αMHD cap per machine | done, approved (OQ-16) | entry 17 |
| Latin hypercube datasets, MAST-U and ITER, 105 cases each | done (version 1) | entry 17 |
| sampled domain plotted against the operating space | done | datasets |
| high-flux cases (K-019) | kept, learned in signed-log space at M3 (OQ-18); not explained | |
| active learning near critical gradients | moved to M3 (OQ-18 a); round 1 in entry 19 | |
| QuaLiKiz driver (Tier A) | deferred (OQ-18 b) | |
| driver interface; Tier B drivers behind the tier flag | interface built; Tier B drivers are stubs that refuse to run without access (OQ-18 c) |
Against the criteria
| criterion | where it stands |
|---|---|
| loads | both datasets load through the library’s loader, which recomputes each file’s hash (100000 of 100000 cases completed, all six fluxes finiteC-071, 100000 of 100000 cases completed, all six fluxes finiteC-072) |
| provenance validates | each file’s SHA-256 matches its provenance record; the generator commit is recorded with a clean tree; a tampered file is refused (unit test) |
| domain plotted | figure; input by input MAST-U 0.72 (triangularity) to 0.94 (q); ITER 0.63 (q) to 0.94 (triangularity shear)C-076. Wider than the measured set jointly (K-018) |
Deviations
- One dataset per machine (owner, 2026-10-02), rather than one sampling space.
- Rejection by αMHD in the sampler, not in the brief: without it, independent draws give pressure gradients far beyond ballooning stability (OQ-16).
- Latin hypercube only at M2. Active learning needs a model of where the thresholds are, so it moved to M3 (OQ-18 a). QuaLiKiz is deferred: its shifted-circle geometry does not represent the MAST-U shapes (OQ-18 b). The Tier B drivers are stubs (OQ-18 c).