Stellarators are fusion devices that use externally generated, three-dimensional magnetic fields to confine hot plasma and offer a promising route to commercial fusion. Their electromagnetic coils pose a challenging design problem: they must reproduce the target magnetic field while satisfying engineering constraints and remaining practical and cost-effective to manufacture. Finding suitable geometries is a highly nonlinear inverse problem that generally has neither a closed-form nor a unique solution. This problem lies at the heart of stellarator design. We introduce CoilStellaration, a dataset and benchmark for learning the mapping between coil geometries from target plasma magnetic fields and engineering requirements. The dataset contains 177,000 optimized filamentary coilsets targeting plasma equilibria from ConStellaration, with sampled engineering requirements and achieved geometric and magnetic-field metrics. It spans multiple coil counts and field-period numbers, supporting studies of both specialized and shared models across configurations. The benchmark asks models to predict coils conditioned on a target equilibrium and requirements on coil-coil and coil-plasma spacing, curvature, and magnetic-field error. We show that magnetic-field accuracy remains the main challenge, with learned models rarely satisfying all requirements simultaneously. However, using learned predictions to initialize conventional coil optimization can accelerate early progress. We release baseline models and evaluation code to support reproducible research on both direct coil generation and learned optimization initializations. By lowering the barrier for machine learning and optimization researchers to engage with stellarator coil design, we aim to accelerate progress toward practical, cost-effective stellarator power plants.
Figures & tables
Figure 1: Left: The target plasma boundary, parameterized as a truncated Fourier surface in θ and ϕ . Center: The generated coilset. Each filamentary coil is modeled as a closed 3D curve using a Fourier series in t , a parameter that runs along the curve. Solid black lines on the stellarator surface trace the equilibrium magnetic field lines, which are tangent to the boundary. Right: Key evaluation metrics. The heatmap displays the normalized magnetic field error across one plasma field period. The diagrams below illustrate the three physical engineering constraints used in this work: minimum coil-to-plasma distance, coil-to-coil spacing, and maximum coil curvature.
Figure 2: End-to-end coilset generation pipeline. A sampled requirement vector parameterizes a Regcoil warm start that is current-rescaled and then refined by a constrained Desc optimization (length vs. quadratic-flux trade-off, subject to coil–coil, coil–plasma, curvature, and total-current constraints). Metrics on both the warm start and the refined coilset are emitted as part of the dataset tuple.
Figure 3: Measured vs. target values for the three geometric constraints. Colour scale is logarithmic. Black line: requirement equals measured. Red line: selection cut threshold.
Figure 4: The filtered dataset spans a broad range of geometric metrics and magnetic-field errors. The pairwise plots show how these quantities vary together. The dataset covers a large range in d~CCmin , d~CPmin , κ~max , l~mean , and field error. The three coilsets are sampled at increasing field error ϵ , reproducing the target plasma boundary with different accuracy. The Poincare plot of the first coilset presents a clean set of flux surface till the plasma boundary, the second one has flux surfaces in the core, but degraded at the boundary, while the third one displays almost no flux surface.
Figure 5: Strict feasibility (%) and mean score S by requirement tier (tight / medium / loose), for the fixed-shape track (top) and the variable-shape track (bottom). Each track shows its own four architectures. Bars are the mean over three training seeds with ±1σ error bars. Across both tracks and all architectures, strict feasibility is 0% on the tight tier and near-zero on the medium tier: nearly all strictly feasible coilsets fall in the loose tier. Strict feasibility uses the benchmark tolerance vi≤1\text{\times}{10}^{-3}$$ .
Table 6
Figure 6: Strict feasibility (left) and mean score S (right), grouped by reference mean normalized field error and field periods. Solid: EnsemblePerCoilMLP, mean ± one standard deviation across three training seeds. Dashed: DESC coilsets. DESC uses the three original geometric bounds; the model uses all four reference-achieved requirements.
Figure 7: Strict-feasibility rate (left) and mean score S (right) against the iteration budget, for the REGCOIL and learned (per-coil ensemble) starts. Each instance uses its dataset metrics as targets; the instances comprise up to 100 per (NFP,NC) pair. Markers show the tested budgets. Shaded bands are instance-level 95% confidence intervals (Wilson for feasibility, Student- t for S ); they do not account for shared boundaries.
On average, 18.4 REGCOIL initializations and 12.2 DESC coilsets
Configuration family
QI-like
QA and QH
QI-like
Number of field periods
1–5
1–6
2–5
Unique coils per half field period
—
1–13
4–8
Appendix
Table 3: Comparison of ConStellaration, QUASR, and CoilStellaration in terms of number of plasma boundaries, coilsets, and their properties. Engineering limits are prescribed requirements, not achieved metric ranges. a is the plasma minor radius, A=R0/a is the aspect ratio, and bCP , bCC , and bκ are sampled requirement thresholds. A dash indicates that a quantity is not applicable.
Figure 8: Fraction of runs meeting each requirement against the iteration budget, for the learned and REGCOIL starts. Shaded bands are 95% Wilson intervals over the instances with results from both starts.
1 iteration
500 iterations
Requirement
Per-coil ensemble
REGCOIL
Per-coil ensemble
REGCOIL
d~CCmin
38%
92%
69%
80%
d~CPmin
58%
10%
72%
52%
κ~max
90%
94%
84%
84%
ϵmax
0%
0%
65%
47%
All
0%
0%
27%
18%
Appendix
Table 4: Fraction of runs meeting each requirement, and all of them (strict feasibility), after 1 and 500 iterations, over the instances with results from both starts. At 500 iterations, every difference except curvature is significant (exact McNemar test, p<10−8 ; curvature p=0.88 ).
Stage
Criterion
Physical / Numerical Rationale
F1
Upstream Convergence
Discard configurations with missing attributes caused by non-convergence in stage-1 (optimization) or stage-2 (forward models).
F2
Optimization Provenance
Require the boundary to originate from active VMEC++ or standard VMEC omnigenity optimization, rejecting unoptimized initial guesses.
F3
Positive Orientation
Enforce ιedge>0 to standardize spatial orientation for downstream control-coil synthesis and neural network ingestion.
F4
Metric Validity
Require a finite Goodman QI residual ( Q=∅ ).
F5
Field Period Bounds
Restrict this release to NFP∈{2,3,4,5} .
F6
Equilibrium Mapping
Exclude boundaries that fail to instantiate or converge to a force-balanced state within the VMEC++ computational framework.
Appendix
Table 5: Summary of the Sequential Target Boundary Filtering Pipeline
Reference Design
d~CCmin (coil-to-coil distance)
d~CPmin (coil-to-plasma distance)
κ~max (coil curvature)
W7-X [ 1 ]
0.53
0.70
1.31
HSX [ 27 ]
0.76
1.18
1.4
NCSX [ 18 ]
0.38
0.67
3.18
Stellaris [ 17 ]
0.51
1.05
2.03
Infinity Two [ 10 ]
0.52
0.97
2.4
Sampled Range
0.3 - 0.7
0.5 - 1.0
1.5 - 5.0
Appendix
Table 6: Reference stellarators from which the requirements sampling range was informed.
NFP
REGCOIL
(%)
DESC
% of DESC
DESC / REGCOIL (%)
2
30 481
11.4
30 234
17.0
99.2
3
104 075
39.0
89 503
50.5
86.0
4
63 609
23.9
33 077
18.7
52.0
5
68 384
25.7
24 527
13.8
35.9
Total
266 549
100.0
177 341
100.0
66.5
Appendix
Table 7: Dataset composition by number of field periods ( NFP ). DESC % of DESC is the fraction of each NFP group within all DESC coilsets. DESC / REGCOIL is the completion rate of the DESC optimization per NFP .
Stage
All
NFP=2
NFP=3
NFP=4
NFP=5
Released samples
266 549
30 481
104 075
63 609
68 384
Successful optimization
177 341
30 234
89 503
33 077
24 527
Full pipeline executed
163 908
30 221
78 906
33 020
21 761
+ Coil-plasma error >−0.5
140 965
26 218
66 237
29 368
19 142
+ Coil-coil error >−0.5
138 954
25 692
65 554
28 949
18 759
+ Curvature error <1.0
138 765
25 614
65 455
28 939
18 757
Appendix
Table 8: Cumulative selection cut flow, overall and by field periods. Retention is relative to the released samples in each column at the time of submission.
Figure 9: Selection effects. (a–c) Probability densities of requested coil–coil clearance, coil–plasma clearance, and maximum coil curvature for released and retained samples. Maximum absolute differences between the corresponding empirical cumulative distribution functions (CDFs) are 0.61, 0.38, and 0.40 percentage points, respectively. (d) Probability density of the minimum normalized magnetic-gradient scale length over unique boundaries; a retained boundary has at least one retained coilset. Panels (a–d) use 20 shared equal-width bins per panel and unit-area normalization for each population. (e) Retained fraction of released samples for each NFP and coil count. (f) Mean per-boundary retention in five scale-length quantile bins per NFP , with equal weight per boundary and points at the mean scale length of each bin.
Metric
Direction
αi
d~CCmin (coil-to-coil distance)
lower bound
0.10
d~CPmin (coil-to-plasma distance)
lower bound
0.10
κ~max (coil curvature)
upper bound
0.10
ϵmax (field error)
upper bound
0.25
Appendix
Table 9: Per-metric tolerances αi in the soft-feasibility score ( Eq. 4 ). A relative violation of αi reduces si to 1/e≈0.37 ; a 2αi violation reduces it to ≈0.14 .
Architecture
sdcc
sdcp
sκ
sϵ
MLP
0.02(v=0.39)
0.35(v=0.11)
0.14(v=0.19)
1×10−5(v=2.8)
ResMLP
0.03(v=0.34)
0.36(v=0.10)
4×10−3(v=0.56)
1×10−5(v=2.9)
EnsembleMLP
0.12(v=0.21)
1.00(v=0)
1.00(v=0)
1×10−4(v=2.2)
EnsembleResMLP
0.09(v=0.25)
1.00(v=0)
1.00(v=0)
2×10−4(v=2.1)
Appendix
Table 10: Median per-metric soft score and (median violation v ) on the fixed-shape track, pooled over the three training seeds per architecture. Soft scores are bounded [0,1] with s=1 at and beyond the threshold and s=1/e at v=α (see Eq. 4 ).
Reference metric
ρ with strict feasibility
ρ with score S
dCCmin
−0.116±0.008
+0.218±0.005
dCPmin
−0.214±0.004
−0.073±0.008
κmax
+0.188±0.003
−0.08±0.02
log10(reference mean normalized field error)
+0.259±0.004
+0.42±0.03
Appendix
Table 11: Spearman correlations between reference-coilset metrics and prediction performance for EnsemblePerCoilMLP on the variable-shape evaluation set. Values are the mean correlation ± one standard deviation across three independently trained seeds.
Figure 10: Mean score Sˉ on the fixed-shape track as the soft-score tolerance is varied: (left) all four tolerances scaled by a common factor c (log axis; dotted line is the nominal c=1 ), and (right) only the field-error tolerance αϵ (dotted line is the nominal 0.25 ). Each point is the mean over three seeds. Field-error-only changes preserve the ranking; uniform scaling can change it.
Uniform scale c
0.25
0.5
1
2
4
τ vs nominal
0.67
1.00
—
0.67
0.33
Field-error αϵ
0.10
0.15
0.25
0.35
0.50
τ vs nominal
1.00
1.00
—
1.00
1.00
Appendix
Table 12: Kendall rank correlation τ of the fixed-shape mean- Sˉ ranking against the nominal- α ranking, as the tolerances are varied. τ=1 means the four-architecture order is unchanged.
Parameter
Value
Description
dh
512
Width of the primary MLP trunk.
L
4
Four hidden stages; separate output projection.
dc
64
Hidden dimension of the specialized currents head.
Dropout
0.2
Dropout rate (exclusive to Residual models).
M
5
Number of members in ensemble variants.
α
1.2
Spectral-scaling exponent for bidirectional flat mapping.
Appendix
Table 13: Hyperparameter Configuration
Parameter
Value
Seeds
0, 1, 2
Training budget
Up to 600,000 updates with 1,000-step warmup.
Batch / selection
Train batch 16, evaluation batch 256, and evaluation every 500 updates.
Stopping Patience
100,000 updates with best checkpoint retained.
Optimizer
AdamW with learning rate schedule 1\times10−3 to 1\times10−5 , and weight decay 1\times10−4 , and global norm clip 1.0.
Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA · Max-Planck-Institut für Plasmaphysik, 85748 Garching, Germany