Using the loss-fraction surrogate#
This tutorial shows how to download our neural surrogate for the loss fraction and use it to predict the loss fraction of a stellarator from the test split, or a new design altogether. The surrogate needs no VMEC run.
Install the package with its models:
pip install "boss-bench[models]"
Now download and load the model using the convenience functions from boss-bench.
download() saves the model in a folder, and
load() loads it:
from boss_bench import models
from boss_bench.models import LossFractionModel
path = models.download("loss_fraction", "BoSS-models/loss_fraction")
model = LossFractionModel.load(path)
A stellarator from the test split#
Download confinement_data_test_split_00000.hdf5 from the
Download Data page into a folder, here BoSS/. We now load a device
from the test data and use the model to predict its loss fraction:
import h5py
import boss_bench as bb
path = bb.list_files("BoSS", "confinement", "test")[0]
with h5py.File(path, "r") as f:
record = bb.load_record(f, 0)
predicted = model.predict(record)
true = bb.loss_fraction(
record["confinement_time"], record["t_max"], record["s_init"], s0=0.1)
print(f"loss fraction: predicted {predicted:.3f}, true {true:.3f}")
loss fraction: predicted 0.264, true 0.252
predict() also takes a batch
of configurations, such as the fields from load().
A new design#
The model reads only the plasma boundary, the pressure and current profiles, the
volume-averaged field strength and the number of field periods, so it can also predict
the loss fraction of a new design that is not in the data. The boundary is given as
boundary_xyz, a (41, 41, 3) array of Cartesian points on one field period of the
boundary surface, indexed \((\phi, \theta, xyz)\). Describe the boundary with a
simsopt SurfaceRZFourier whose quadrature points
are 41 toroidal angles over one field period and 41 poloidal angles, uniformly spaced
and without the endpoint. As shown below, boundary_xyz = surface.gamma() is then the
correct input for the model. The profiles are evaluated at 15 radial points linearly
spaced from \(s = 10^{-4}\) to 1, the grid s1d of the data.
Here the design is a rotating ellipse at the scale of ARIES-CS: three field periods,
a major radius of 7.75 m, a minor radius of about 1.7 m and a volume-averaged field of
5.86 T, in vacuum, so with zero pressure and current (pip install simsopt):
import numpy as np
from simsopt.geo import SurfaceRZFourier
nfp = 3
surface = SurfaceRZFourier(
nfp=nfp, stellsym=True, mpol=1, ntor=1,
quadpoints_phi=np.linspace(0, 1 / nfp, 41, endpoint=False),
quadpoints_theta=np.linspace(0, 1, 41, endpoint=False))
surface.set_rc(0, 0, 7.75) # major radius [m]
surface.set_rc(1, 0, 1.7) # minor radius [m]
surface.set_zs(1, 0, 1.7)
surface.set_rc(1, 1, 0.5) # the rotating ellipse
surface.set_zs(1, 1, -0.5)
design = {
"boundary_xyz": surface.gamma(), # (41, 41, 3) [m]
"pressure": np.zeros(15), # [Pa]
"toroidal_current": np.zeros(15), # [A]
"volume_avg_modB": 5.86, # [T]
"nfp": nfp,
}
print(f"predicted loss fraction: {model.predict(design):.3f}")
predicted loss fraction: 0.272
A plain rotating ellipse is far from an optimized stellarator: the model predicts that it loses about a quarter of the alpha particles born on \(s = 0.1\).
The inputs are described in LossFractionModel.
Calling the model on PyTorch tensors instead of
predict() gives the prediction
with gradients, for example with respect to the boundary, for gradient-based
optimization.