CoilStellaration: A Dataset and Benchmark for Engineering-Aware Stellarator Coilset Generation

Jan-Hendrik Ewers, Santiago A. Cadena, Nicolò Foppiani, Andrea MerloarXiv 2026 · arXiv

TL;DR

CoilStellaration is a massive new dataset and benchmark featuring 177,000 optimized stellarator coilsets designed to bridge the gap between plasma physics and engineering feasibility. While direct machine learning (ML) generation of "buildable" coils remains a significant challenge, the study demonstrates that ML models can provide powerful "warm starts" that significantly accelerate traditional optimization workflows.

What Problem Does This Paper Address?

In the race for commercial fusion, stellarators offer a compelling alternative to the more common tokamak. While tokamaks rely on internal plasma currents for stability, stellarators use complex, 3D external coils to create the necessary magnetic "cage." This makes stellarators inherently stable but creates a monumental engineering headache: coil design.

Designing these coils is a "nonlinear inverse problem." Engineers must find a set of complex 3D curves that produce a precise magnetic field while satisfying strict manufacturing constraints:

  1. Coil-to-plasma spacing: Coils can't be too close to the hot plasma.
  2. Coil-to-coil spacing: There must be room for structural supports and assembly.
  3. Curvature: Superconducting cables can only bend so much before they break or lose performance.

Traditional optimization is computationally expensive and slow. This paper introduces a data-driven framework to see if ML can learn to map desired plasma shapes and engineering requirements directly to viable coil geometries.

Methods

The authors developed a multi-stage pipeline to generate the dataset, as illustrated in the figure below:

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.
  1. Equilibria Selection: They started with plasma boundaries from the ConStellaration dataset.
  2. Requirement Sampling: They sampled unique sets of engineering constraints (spacing, curvature) based on real-world reactor designs.
  3. Two-Stage Optimization: They used a fast initial method (REGCOIL) to get a "warm start," followed by a rigorous refinement using the DESC framework to enforce constraints.
  4. ML Benchmarking: They tested several architectures (MLPs, ResMLPs, and Ensembles) on two tracks: a fixed-shape track (specific coil topology) and a variable-shape track (handling different numbers of coils and field periods).

Datasets

The primary contribution is the CoilStellaration dataset, which contains:

  • 177,341 optimized DESC coilsets.
  • 14,496 unique plasma boundaries.
  • A wide range of configurations, covering 2 to 5 field periods and 4 to 8 unique coils per half-period.

This dataset is significantly more diverse in its engineering requirements than previous repositories like QUASR, which often focus on a narrower set of design parameters.

Key Insights and Ideas

  • The "Unforgiving" Metric: The authors use a geometric mean of four sub-scores to evaluate models. In engineering, if a coilset fails even one constraint (e.g., it's too close to the plasma), the entire design is useless. This metric reflects that "all-or-nothing" reality.
  • Magnetic Field Accuracy is the Bottleneck: While models can learn to satisfy geometric constraints (like spacing), reproducing the target magnetic field with high precision remains the hardest part of the problem.
  • ML as a Catalyst, Not a Replacement: Current ML models rarely produce a perfect, "strictly feasible" coilset in a single forward pass. However, they are excellent at getting "close."
  • Warm-Starting Optimization: The most immediate value of these models is as an initialization for traditional optimizers. A "learned start" reaches better solutions much faster than standard initialization methods.

Key Results

The benchmark results show that the problem is far from solved, but the progress is measurable:

  • Low Direct Feasibility: The best baseline model achieved a "strict feasibility" rate of only 4.4% on the benchmark. Most successes occurred when engineering requirements were "loose."
  • Superior Warm Starts: When used to initialize a DESC optimization, the ML-predicted coilsets outperformed the standard REGCOIL initialization. As shown in Figure 7, the learned start achieved a higher feasibility rate in 200 iterations than the standard method did in 500.
Comparison of strict feasibility and mean score S between REGCOIL and learned warm starts over optimization iterations.

Limitations and Future Work

  • Filamentary Approximation: The dataset models coils as infinitely thin filaments. In reality, coils have a finite thickness ("winding packs") that affects both the magnetic field and the physical space they occupy.
  • Limited Constraints: While a step forward, the benchmark doesn't yet include all real-world constraints, such as forces between coils or complex support structures.
  • Loss Function Refinement: Future work could involve training models with loss functions that directly account for the magnetic field error, rather than just the geometric coordinates of the coils.

By releasing this dataset and code, the authors hope to invite the broader ML community to help solve one of the most difficult engineering hurdles in the path to clean fusion energy.

Summary updated

Abstract

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.
BibTeX
@misc{ewers2026coilstellaration,
  author = {Jan-Hendrik Ewers and Santiago A. Cadena and Nicolò Foppiani and Andrea Merlo},
  title = {{CoilStellaration: A Dataset and Benchmark for Engineering-Aware Stellarator Coilset Generation}},
  year = {2026},
  eprint = {2611.12689},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12689},
}
Figure