TGLF datasets for MAST-U and ITER, version 1
The surrogates of M3 learn TGLF’s fluxes from examples (entry 16). The owner set one dataset per machine, MAST-U and ITER, so that either can be used alone or both pooled. Each dataset is a set of TGLF input points and the fluxes TGLF returns for them.
How the points were chosen
Each TGLF input is a local description of the plasma at one radius. We split the inputs into two groups. The geometry inputs (major radius, elongation, triangularity and their radial derivatives, q, magnetic shear) were measured, not guessed: we computed them the way TORAX computes them, on the flux surfaces of real equilibria, and took each range from the 1st to 99th percentile. For MAST-U the equilibria are 42 FreeGSNKE free-boundary solutions on the MAST-U machine description. For ITER they are TORAX’s ITER hybrid equilibrium plus the q profiles of TORAX’s current ramp-up example. The kinetic inputs (temperature and density gradients, temperature ratio, beta, collisionality, effective charge, flow shear) were proposed by the agent and approved by the owner per machine (OQ-16). The generator refuses to write a dataset for a region until its kinetic inputs are marked approved.
The inputs are drawn by Latin hypercube sampling. Drawn independently, high beta and steep gradients combine into pressure gradients far beyond ballooning stability, where TGLF returned fluxes of 104 to 105 gyro-Bohm units in the timing run. Points are therefore rejected above a fixed normalised pressure gradient αMHD, set per machine for version 1; a cap that depends on the magnetic shear is to follow (OQ-16). The cap keeps about half the draws and shifts the kept set towards low q and low beta.
What was produced and checked
Two cluster jobs, one per machine, each of 105 cases. Each dataset is written once, made read-only, and stored with a provenance record that binds the file’s hash, the sampling space and the measured geometry. Both are also stored on the run tracker (Simvue) with their provenance, and the uploaded files match the local hashes.
| quantity | MAST-U and ITER |
|---|---|
| cases completed, fluxes finite | 100000 of 100000 cases completed, all six fluxes finiteC-071; 100000 of 100000 cases completed, all six fluxes finiteC-072 |
| draws kept under the cap | MAST-U 0.490 kept (220936 drawn), ITER 0.425 (258597 drawn); kept mean beta_e 23-25% below the box mean on both; kept median q 1.90 (MAST-U) and 1.73 (ITER, sampled log-uniform over 1.02-11.1)C-073 |
| cost | 91.6 core-h (MAST-U) and 92.4 core-h (ITER); 3.3 s per case mean; 35 min wall per jobC-074 |
| cases with Qe + Qi above 104 gyro-Bohm units | MAST-U 0.283 above 10^4 and 0.122 above 10^5; ITER 0.097 above 10^4 and 0.0065 above 10^5C-075 |
The brief’s acceptance test for M2 includes plotting the sampled domain against the intended operating space. The plot is in the technical section. It shows that the geometry inputs, drawn independently, cover combinations that no measured equilibrium has, for example high q at low shear on MAST-U (MAST-U 0.72 (triangularity) to 0.94 (q); ITER 0.63 (q) to 0.94 (triangularity shear)C-076; ceiling K-018).
The last row is not explained yet. Even under the cap, a large fraction of cases have heat fluxes several orders of magnitude above typical values, mostly at steep electron temperature gradients. We have not established whether these are physical, for example electromagnetic modes at high beta, or come from how the inputs are constructed. The cases are kept as computed, and the question is recorded as ceiling K-019, to be resolved before training at M3.
Where this stands
M2: datasets exist, load and validate. Open before M3: K-016 (ITER v1 has one plasma shape), K-017 (fixed cap), K-018 (independent geometry sampling), K-019 (high-flux cases).
Technical details → Datasets. Decisions → OQ-15, OQ-16.
Technical details → Datasets › sampling space · Datasets › datasets v1
Decisions → OQ-15 · OQ-16