🔬 physics

Augmented Lagrangian methods produce cutting-edge magnetic coils for stellarator fusion reactors

This paper introduces an augmented Lagrangian approach to overcome the challenges of nonconvex optimization in stellarator design, successfully generating superior, Pareto-optimal magnetic coil configurations for five diverse stellarator models.

Original authors: Pedro F. Gil, Weiping Li, Julianne Stratton, Alan A. Kaptanoglu, Eve V. Stenson

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pedro F. Gil, Weiping Li, Julianne Stratton, Alan A. Kaptanoglu, Eve V. Stenson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to build a cage for a star. That is the dream behind fusion energy: trapping super-hot plasma in a magnetic bottle to generate limitless, clean power. The most promising design for this cage is called a "stellarator." Unlike a simpler, donut-shaped version called a tokamak, a stellarator twists its magnetic field into a complex, three-dimensional spiral. This twist is crucial because it keeps the plasma stable without needing a massive electric current that could cause the whole thing to wobble apart. However, there's a catch: to create this twisted magnetic field, you need to build a set of external wire coils that are shaped like bizarre, knotted ribbons floating in space.

The problem is that designing these coils is like trying to solve a puzzle where the pieces keep changing shape, and the rules are incredibly strict. You need the coils to be close enough to the plasma to hold it tight, but far enough away so they don't melt or touch. They need to be smooth enough to manufacture, but complex enough to create the right magnetic twist. If the coils are even slightly off, the magnetic cage fails, and the star escapes. For decades, scientists have struggled to find the perfect coil shapes, often spending years and millions of dollars just to figure out how to build them, only to find that the math says the design is impossible to build with current technology.

This paper introduces a new, smarter way to solve that puzzle. The authors, a team of physicists and mathematicians, developed a method called the "Augmented Lagrangian" approach to design these magnetic coils. Think of the old way of designing coils like trying to bake a cake by guessing the amount of sugar, flour, and eggs. You mix a batch, taste it, and if it's too sweet, you guess how much less sugar to use next time. You might have to bake dozens of cakes, tweaking the recipe manually, just to get something edible. In the world of stellarators, this meant scientists had to manually adjust "weights" (mathematical knobs) to balance the need for a strong magnetic field against the need for the coils to be buildable. Often, they would get stuck, or the resulting coils would be too wiggly to manufacture, or they would cost a fortune.

The new method described in the paper is more like having an AI chef that automatically adjusts the recipe while you bake. Instead of guessing the right balance, the computer uses a mathematical system that constantly checks the "rules" (like "coils must not touch" or "magnetic field must be perfect") and automatically shifts the focus to fix whatever is broken. If the coils are too close, the system tightens the penalty on distance; if the magnetic field is weak, it focuses on accuracy. It does this without a human needing to constantly tweak the settings.

Using this new tool, the researchers tested it on five different stellarator designs, ranging from theoretical concepts to real, existing machines like the Wendelstein 7-X in Germany and the HSX in the US. In every case, the new method found coil designs that were better than the ones previously published. For example, for a theoretical design called "Stellaris," they found a version with fewer coils that still held the plasma perfectly, leaving more space for maintenance and diagnostics. For the real-world Wendelstein 7-X, they found new coil shapes that were smoother and easier to build, potentially saving years of construction time and reducing the risk of manufacturing errors.

The paper shows that this new approach can navigate the messy, complicated math of coil design much more efficiently than before. It doesn't just find a solution; it finds solutions that sit on the "Pareto frontier," meaning you can't improve one aspect (like making the magnetic field more accurate) without making another aspect worse (like making the coils too expensive or hard to build). By automating the balancing act, the authors suggest that we can now design stellarators that are not just theoretically possible, but actually practical to build, bringing the dream of fusion power one step closer to reality.

Technical Summary: Augmented Lagrangian Methods for Stellarator Coil Optimization

Problem Statement
The design of magnetic coils for stellarator fusion reactors represents a critical bottleneck in realizing controlled nuclear fusion. While stellarators offer steady-state operation without the need for induced plasma currents (unlike tokamaks), their complex three-dimensional magnetic fields require intricate external coil systems. The optimization of these coils is an ill-posed, highly non-convex inverse problem. Traditional design workflows, which separate plasma boundary optimization (Stage I) from coil shape optimization (Stage II), struggle with this complexity.

Current methodologies typically rely on transforming the constrained optimization problem into an unconstrained multi-objective problem by assigning manual weights to various objectives and constraints (e.g., field accuracy, coil curvature, inter-coil distance). This approach suffers from several limitations:

Manual Tuning: The weights lack physical significance and require extensive, computationally expensive manual tuning to find feasible solutions.
Local Minima: The non-convex landscape often traps optimization algorithms in suboptimal configurations.
Engineering Infeasibility: Many generated coil sets are impractical due to excessive curvature, interlinking, or forces exceeding material tolerances, leading to construction delays and cost overruns in projects like Wendelstein 7-X and the cancellation of NCSX.

Methodology
The authors propose an Augmented Lagrangian (AL) method implemented within the SIMSOPT optimization suite to address these challenges. Unlike traditional weighted-sum approaches, the AL method treats constraints dynamically.

Formalism: The optimization minimizes a Lagrangian function $L_A(x, \lambda, \mu) = f(x) - \lambda^\top c(x) + \frac{1}{2} \|\sqrt{\mu} \circ c(x)\|_2^2$ , where $f(x)$ is the objective, $c(x)$ represents constraints, $\lambda$ are Lagrange multipliers, and $\mu$ are penalty parameters.
Dynamic Updates: Instead of fixed user-defined weights, the Lagrange multipliers ( $\lambda$ ) and penalty parameters ( $\mu$ ) are automatically updated throughout the optimization process based on constraint violations. This eliminates the need for manual weight tuning.
Constraint Handling: The method treats the squared flux error (a measure of magnetic field deviation) as a constraint rather than the primary objective. By setting a threshold (e.g., $f_{SF} \lesssim 10^{-6}$ ), the algorithm prioritizes satisfying engineering constraints once the field accuracy is sufficient, rather than chasing a theoretically perfect but practically impossible zero-error solution.
Engineering Constraints: The framework incorporates standard metrics for coil complexity, including minimum coil-to-coil and coil-to-surface distances, total coil length, curvature limits, mean squared curvature, Gauss linking numbers (to prevent interlinking), and electromagnetic forces.
Optimization Algorithm: The inner minimization step utilizes the L-BFGS-B quasi-Newton method. Gradients are computed using a combination of analytic derivatives and automatic differentiation via vector-Jacobian products.

Key Contributions and Results
The paper demonstrates the effectiveness of the AL method by generating coil sets for five distinct stellarator configurations, ranging from theoretical concepts to fully constructed devices. In all cases, the method produced Pareto-optimal solutions that outperform or match previously published sets without manual weight tuning.

Quasi-Axisymmetric (QA) Stellarator:
- The method mapped the Pareto front between coil forces and field accuracy for a Landreman&Paul QA configuration.
- It generated a 3-coil per half-field period solution that outperformed a reference 4-coil solution (Wechsung et al.) in field accuracy and coil-coil distance, despite the expectation that fewer coils would increase ripple.
- It successfully traced new Pareto boundaries by relaxing total length constraints, achieving field accuracies up to an order of magnitude better than previous scans.
Quasi-Helical Symmetric (QH) Stellarator:
- Applied to a Landreman&Paul QH configuration, the method produced a 4-coil solution that matched the physics performance of a reference 5-coil solution (Wiedman et al.) while improving engineering metrics (e.g., reduced coil length, increased spacing).
- A third configuration optimized purely for engineering properties showed a 38% increase in coil-to-surface distance and a 10% reduction in HTS material cost, albeit with higher quasisymmetric error, demonstrating the ability to trade physics fidelity for manufacturability.
Stellaris (SQuID) Configuration:
- For the Proxima Fusion Stellaris concept, the method found alternatives with 5 and 6 coils per half-field period.
- These solutions improved magnetic field error by 25% and increased inter-coil spacing by 13 cm compared to the original design, while reducing peak forces on HTS turns by 29%.
- A configuration optimized for engineering properties achieved a 6% reduction in estimated HTS cost and significantly simpler geometries.
Wendelstein 7-X (W7-X):
- The method generated alternative coil sets for the existing W7-X reactor.
- Configuration #1: Reduced the coil count from 50 to 40 (5 to 4 per half-field period) while improving field accuracy and reducing total coil length.
- Configuration #2: Achieved a 20% reduction in electromagnetic forces with only a 5% increase in total coil length.
- Configuration #3: Dramatically improved field error and curvature with negligible changes to length or forces.
- All alternatives exhibited lower curvature and smoother geometries than the original design.
Helically Symmetric eXperiment (HSX):
- For the existing HSX device, the method produced a 4-coil solution that reduced total coil length by 25% and curvature by 45% compared to the original 6-coil design, while maintaining reasonable field accuracy. This highlights the potential for simplifying manufacturing in research-scale devices.

Significance and Claims
The paper claims that the Augmented Lagrangian approach fundamentally advances the state-of-the-art in stellarator coil design by:

Automating Weight Tuning: Removing the reliance on heuristic, manual weight adjustments, which are a major source of inefficiency and suboptimality in current workflows.
Robustness: Successfully navigating the non-convex optimization landscape to find feasible, Pareto-optimal solutions across diverse magnetic configurations (QA, QH, QI).
Versatility: Demonstrating applicability to both conceptual designs and existing, fully manufactured reactors (W7-X, HSX), proving the method can improve upon established engineering solutions.
Efficiency: Achieving these results with significantly less computational overhead than the massive parameter scans (e.g., 8,500 runs) required by previous stochastic or weighted-sum methods.

The authors conclude that this methodology provides a computationally efficient and robust pathway for generating coil sets that balance plasma physics requirements with the stringent engineering constraints necessary for reactor-scale integration.

Technical Summary: Augmented Lagrangian Methods for Stellarator Coil Optimization

More like this