Sparse Coil Optimization in Fusion Coil Design

• Sparse coil optimization is a method that uses inverse problem formulations and sparsity-promoting regularizers to select a minimal set of coils for stellarator design.
• It leverages algorithms such as LASSO, relax-and-split, OtD, and MIQP to balance magnetic field accuracy with manufacturability under engineering constraints.
• Benchmark results show that MIQP reduces mean magnetic field errors by up to 20% and LASSO cuts total current by up to 12% at fixed accuracy, enhancing plasma confinement.

Sparse coil optimization refers to the family of algorithmic strategies for designing magnetic field coils that minimize coil count, improve manufacturability, and reduce engineering complexity while retaining precise magnetic field control, especially in the context of stellarator devices for fusion energy. Contemporary approaches utilize sparse regression, differentiable proxies, and convex or combinatorial formulations to reconcile the physical requirements of plasma confinement with practical and economic constraints on coil complexity (Wu et al., 11 Feb 2025, Fu et al., 17 Oct 2025).

1. Mathematical Formulation of Sparse Coil Optimization

The sparse coil optimization problem is typically cast as an inverse problem involving the Biot–Savart law, wherein the target is to approximate a desired magnetic field (often quantified by the normal component $B_n$ on the plasma boundary) using a minimal, strategically-selected subset of potential coil locations or current distributions. The canonical formulation is: $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$ Here, $I = (i_1,\ldots,i_m)$ denotes the currents in $m$ candidate coils, $A\in\mathbb{R}^{N\times m}$ is the discretized Biot–Savart matrix mapping currents to the plasma-surface normal field, $b$ is the target normal field after accounting for background contributions, and $R(I)$ encodes a regularization promoting sparsity (e.g., $\ell_1$, $\ell_0$ norms) or coil complexity. Engineering bounds on current magnitudes are explicit constraints.

In winding-surface methods, coil sets are parameterized as surface-current densities $K = \hat{n}\times\nabla\Phi'$ on a prescribed surface $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$0, with sparsity proxies including the maximum sheet current density and the squared norm of the current potential: $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$1 These proxies permit a continuous quantification of coil “sparsity” or engineering cost during high-level plasma optimization (Fu et al., 17 Oct 2025).

2. Optimization Algorithms and Regularization Approaches

Multiple algorithmic paradigms are deployed to enforce sparsity in current distributions or coil proxies:

• LASSO ($$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$2 Penalty):

$$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$3

Promotes entrywise sparsity, allowing convex solvers (coordinate descent or proximal-gradient) to trace full error-sparsity Pareto fronts efficiently.

• Relax-and-Split ($$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$4 Penalty):

$$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$5

Using auxiliary variables and alternating updates (e.g., proximal hard-thresholding), this heuristic attacks the combinatorial nature of support selection with reduced computational complexity compared to direct MIQP solvers.

• Optimize-then-Delete (OtD):

Sequentially removes coils with the lowest current magnitude after each re-optimization step, producing a hierarchy of increasingly sparse solutions. Randomized variants select coils for deletion probabilistically, exploring distinct sparsity patterns.

• Mixed-Integer Quadratic Programming (MIQP):

Binary variables encode coil activity, enforcing explicit cardinality constraints. Commercial MIQP solvers (e.g., Gurobi) generate globally optimal or near-optimal coil patterns for fixed removal targets at nontrivial computational cost.

• Masked Re-optimization:

For any predetermined sparsity pattern, re-solving the least-squares system within active coil indices yields minimum-norm current sets for given support (Wu et al., 11 Feb 2025).

Coil complexity in winding-surface models is tackled by formulating the subproblem as a nonconvex QCQP, with augmented Lagrangian methods enabling high-efficiency optimization and adjoint-based differentiation for gradient-based outer-loop solvers (Fu et al., 17 Oct 2025).

3. Pareto Front Construction and Quantitative Trade-Offs

Sparse coil optimization inherently involves a trade-off: reducing coil count generally worsens magnetic-field accuracy. Pareto fronts are constructed by plotting key error metrics—mean and maximum plasma-surface $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$6—versus sparsity (e.g., number of nonzero currents or dipole-sheet density). Quantitative comparisons show:

Method	Mean $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$7 (T)	Max $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$8 (T)	Total Current (MA)
OtD	0.0127	0.088	165.7
LASSO	0.0134	0.148	144.7
MIQP	0.0102	0.076	159.2

MIQP consistently generates fronts with the lowest mean field error for a given coil count, while LASSO minimizes total current at fixed error. OtD patterns degrade at high sparsity.

In winding-surface quasi-single-stage optimization, using differentiable coil proxies leads to 20–50% reductions in coil complexity measures (peak and total dipole count) at fixed plasma physics targets relative to reference equilibria (Fu et al., 17 Oct 2025).

4. Sensitivity and Robustness to Manufacturing Errors

Sparse solutions may exhibit enhanced or reduced sensitivity to coil geometry perturbations. Misalignments are modeled by randomly translating and rotating coil centroids or modifying winding-surface parameters, and the mean increase in $$\min_{I\in\mathbb{R}^m} \; \frac{1}{2}\|A I - b\|_2^2 + \lambda\, R(I) \quad\text{s.t.}\quad \|I\|_\infty \leq i_\mathrm{max}$$9 is measured across ensembles. Higher sparsity (fewer coils) generally reduces aggregate sensitivity since fewer elements are disturbed, and LASSO-based solutions demonstrate slightly more robustness at moderate sparsity. There is a direct connection between robustness to misalignment and regularization: robust regression formulations (weighted LASSO) naturally arise in this context, and their optimal solutions approximate the best-perturbation responses empirically (Wu et al., 11 Feb 2025).

5. Integration with Equilibrium Optimization and Proxy Methods

Modern approaches increasingly couple coil sparsity/complexity proxies directly into plasma equilibrium design:

• Quasi-single-stage optimization: Minimizes a composite cost—plasma physics objectives plus a differentiable coil-complexity proxy—subject to plasma and engineering constraints, where the proxy can be the maximum dipole-sheet density or the total dipole count on a winding surface.
• Adjoint-based differentiation: The gradient of the coil complexity with respect to plasma boundary parameters is computed using implicit differentiation on the augmented Lagrangian stationarity conditions, with all Jacobians approximated by automatic differentiation except the inner linear solve (Fu et al., 17 Oct 2025).
• Ill-posedness avoidance: Fixing the winding surface and solving for current potentials ensures convexity and regularity, avoiding degeneracies inherent to filamentary current models.
• Self-intersection–free, differentiable winding-surface construction: Algorithmic schemes impose geometric regularity needed for differentiable, well-posed outer-loop optimization, and are integrated into open-source toolchains such as QUADCOIL.

6. Benchmark Results and Practical Implications

Benchmark studies on the Eos stellarator and MUSE-like configurations illustrate practical impact:

• For planar shaping coils on Eos, MIQP achieves up to 20% reductions in mean $I = (i_1,\ldots,i_m)$0 error at a fixed coil count compared to heuristics, and LASSO reduces total current for a given accuracy by up to 12%.
• In MUSE-based configurations using differentiable coil proxies, up to 50% reductions in total dipole count are achieved at equivalent confinement and quasi-symmetry.
• All methods are sensitive to the selection of proxy, smoothing, or regularization parameters, with marked differences in resulting coil patterns.

These results suggest that sparse optimization enables selection among candidate coil sets that optimally balance plasma performance, manufacturability, and robustness to engineering tolerances (Wu et al., 11 Feb 2025, Fu et al., 17 Oct 2025).

7. Future Directions

Ongoing and prospective research directions include:

• Extending adjoint systems to efficiently handle extremely large sets of constraints.
• Transitioning from quasi-single-stage to fully single-stage optimization, wherein plasma and coil degrees of freedom are jointly optimized under convex coil objectives, obviating the need for nested subproblem solves.
• Expanding robust optimization methodologies for coil design under broader perturbation ensembles.
• Developing custom QCQP solvers or KKT-aware stationarity mapping for improved adjoint accuracy at scale.
• Further exploration of winding-surface geometry generation and nonconvex, physically-motivated regularizers for coil complexity.

These advances are positioned to fundamentally improve the engineering feasibility and economic viability of optimized fusion devices while maintaining tight control over magnetic equilibrium properties (Wu et al., 11 Feb 2025, Fu et al., 17 Oct 2025).