| active learning |
choosing the next training points with the help of the current model, instead of sampling them in advance |
| ASCOT5 |
open-source code that follows fast particles (beam ions, fusion alphas) through the plasma to work out where they deposit heat and drive current |
| Aurora |
open-source code for impurity transport and radiation |
| benchmark |
a published simulation that a new code must reproduce before its own results are believed |
| Bohm/gyro-Bohm |
a simple empirical model of turbulent transport; used here as the trusted fallback |
| checkpointing |
recomputing parts of a calculation instead of storing them, to keep the memory for derivatives bounded |
| claim ID |
[C-012](/Blog/tokamak-toolkit/claims.qmd#c-012) etc.: the row in the claims register that a quoted number comes from |
| conformal prediction |
a statistical method that turns a model’s errors on held-out data into error bars with a guaranteed coverage rate |
| coupling loop |
the part of the simulation that makes the physics modules agree with each other at each time step |
| diverted plasma |
a plasma whose outermost surface has an X-point, a magnetic null that funnels exhaust onto dedicated targets (the divertor) |
| ensemble |
several copies of a neural network trained independently; their spread measures how uncertain the prediction is |
| EQDSK |
a standard file format for a tokamak equilibrium: flux on an R–Z grid, the boundary, and pressure and current profiles |
| equilibrium |
the shape and internal magnetic structure the plasma settles into |
| fallback |
the trusted physics model used wherever the guard rejects the surrogate |
| fallback fraction |
share of (radius, time step) points served by the fallback; reported for every run |
| FF′ |
the product of the poloidal current function F and its derivative with respect to flux; with the pressure gradient it sets the current density in a Grad–Shafranov equilibrium |
| finite differences |
estimating a derivative by nudging an input and watching the output; used to check exact derivatives |
| fixed point |
a state that maps onto itself; each time step solves for one |
| flat-top |
the steady phase of a plasma discharge, after the current has ramped up |
| FreeGSNKE |
open-source free-boundary equilibrium solver (LGPL-3.0) |
| Grad–Shafranov equation |
the force-balance equation an axisymmetric tokamak equilibrium satisfies |
| growth rate (γ) |
how fast a linear instability grows: its energy rises as e2γt |
| guard |
the per-point check that decides whether a surrogate’s answer may be used |
| gyro-Bohm unit (gB) |
the normalisation TGLF reports its fluxes in, built from the local temperature, density, magnetic field and minor radius |
| IMAS / IDS |
the fusion community’s standard data model; an IDS is one record type in it (e.g. equilibrium, core_profiles) |
| ITER |
the international tokamak under construction in France; its “hybrid scenario” is the benchmark here |
| JAX |
numerical library that compiles and differentiates array programs |
| jit |
just-in-time compilation: turning a Python function into one compiled program |
| JOREK |
open-source nonlinear extended-MHD code (LGPL-3.0), used here for edge-stability reference runs |
| MAST-U |
the UK’s spherical tokamak at Culham |
| MHD |
magnetohydrodynamics: the plasma treated as one electrically conducting fluid |
| Newton-Raphson |
an iterative solver that uses derivatives; converges very fast near the answer |
| pedestal |
the steep-gradient region just inside the plasma edge in high-confinement mode |
| peeling-ballooning modes |
edge instabilities driven by the pedestal’s current (peeling) and pressure gradient (ballooning) |
| Picard iteration |
the simplest iterative solver: guess, recompute, repeat |
| provenance |
a record, stored next to every artifact, of the code version, configuration and inputs that produced it |
| q (safety factor) |
how many times a field line goes round the torus the long way per turn the short way |
| QLKNN |
a neural network trained to reproduce the QuaLiKiz turbulent transport code; part of TORAX |
| ramp-up |
the phase in which the plasma current is raised from zero to its target |
| residual |
how far an iterative solver is from satisfying its equations |
| ρ (rho) |
normalised radius: 0 at the centre, 1 at the edge (strictly, based on toroidal magnetic flux) |
| sawtooth |
a periodic internal collapse of the core temperature |
| separatrix |
the plasma’s last closed flux surface; in a diverted plasma it passes through the X-point |
| spherical tokamak |
a tokamak with a very small central hole relative to its size, like MAST-U |
| stiff |
a system where small input changes cause large output changes; hard for simple solvers |
| surrogate |
a fast model (here a neural network) trained to reproduce a slow one |
| TGLF |
quasilinear model of plasma turbulence from General Atomics’ GACODE suite (Apache-2.0); the source of the training data for the transport surrogates |
| Tier A / Tier B |
open-source codes the toolkit may depend on / UKAEA-access-gated codes it may only learn from |
| TORAX |
open-source tokamak transport simulator in JAX, from Google DeepMind (Apache-2.0) |
| toroidal mode number (n) |
how many times a perturbation repeats around the torus |
| transport |
the leakage of heat and particles across the magnetic surfaces |