Runs the EPEDNN pedestal model.
Both models are shipped in data/ and are loaded with EPEDNN.loadmodelonce(<filename>):
| model | inputs | outputs |
|---|---|---|
EPED1NNmodel.bson |
10: a, betan, bt, delta, ip, kappa, m, neped, r, zeffped |
18: pedestal pressure and width, for 3 diamagnetic models × 3 solutions |
EPED1NNensemble.bson |
12: the same, plus nesep_ratio (ne_sep/ne_ped) and tesep |
1: pedestal pressure (dmagH diamagnetic model, first solution), with an uncertainty |
Both are a power law fitted in log space plus a neural-network correction on top of it, so that predictions revert to the physics-shaped power-law trend where the training data runs out.
input = EPEDNN.InputEPED()
input.a, input.betan, input.bt, input.delta, input.ip = 2.0, 2.0, 5.3, 0.49, 15.0
input.kappa, input.m, input.neped, input.r, input.zeffped = 1.85, 2.5, 7.0, 6.2, 1.5
input.nesep_ratio, input.tesep = 0.25, 75.0 # only used by the ensemble
epedmod = EPEDNN.loadmodelonce("EPED1NNmodel.bson")
solution = epedmod(input) # pedestal pressure [MPa] and width [ψ_N] of each solution
solution.pressure.GH.H, solution.width.GH.H
ensemble = EPEDNN.loadmodelonce("EPED1NNensemble.bson")
ensemble(input) # (height, sigma, sigma_frac, extrapolation, sigma_frac_combined, in_distribution)The same three calls work for either model, so that a code can switch between them without knowing which one it is holding:
sol = EPEDNN.run_epednn(pedmodel, input)
pressure, uncertainty = EPEDNN.pedestal_height(pedmodel, input, sol) # [MPa], and as a fraction of it
width = EPEDNN.pedestal_width(pedmodel, sol, βpol_ped) # [ψ_N] 1/2 widthβpol_ped is the poloidal beta of the pedestal pressure (IMAS.pedestal_poloidal_beta), and is only
used by the models that predict the height alone.
EPED1NNensemble is a deep ensemble: several networks trained independently on the same data, whose
disagreement measures the uncertainty of the prediction. ensemble_uncertainty combines it with the
distance from the training bounds (extrapolation_distance) into an in_distribution flag. The two
are complementary: the ensemble spread catches points that are within the bounds of each individual
input but away from the (thin) manifold where the training data actually lives, while the distance
catches points far outside of the bounds, where all the members fall back onto the same power law and
so agree with each other for the wrong reason.
The shipped ensemble predicts the pedestal height only: the width of an EPED solution follows the
analytic EPED1 width law w_ped = 0.076·√βₚ,ped (pedestal_width), so there is nothing left for a
network to learn. It was trained on the multi-machine EPED database used for EPED1NNmodel plus new
ITER-scale EPED runs, which extend the training set to low pedestal density and to separatrix
conditions (nesep_ratio, tesep) that the original database sampled at a single value.
scripts/pytorch_ensemble_to_bson.jl packages a trained ensemble into the data/*.bson shipped here.
For more details, see the online documentation.
