A flexible and differentiable coil proxy for stellarator equilibrium optimization

This paper introduces a flexible, differentiable coil complexity proxy based on the fast QUADCOIL code to enable quasi-single-stage stellarator optimization, effectively balancing plasma performance and coil metrics while successfully reducing magnet counts for MUSE and coil forces for ARIES-CS.

The Gist

Imagine you are an architect trying to build a futuristic, floating city inside a giant, invisible magnetic bubble. This bubble, called a stellarator, is designed to hold super-hot plasma (like the stuff in the sun) to generate clean, limitless energy.

The biggest challenge isn't making the bubble; it's building the magnetic coils (the giant electromagnets) that hold it together.

The Old Way: The "Two-Step Dance"

For a long time, engineers designed these stars in two separate steps:

1. Step 1: They designed the perfect shape of the magnetic bubble to hold the plasma. They made it as efficient and stable as possible, ignoring how hard it would be to build the magnets.
2. Step 2: They tried to build magnets to match that perfect bubble.

The Problem: Often, the "perfect" bubble from Step 1 required magnets that were impossibly complex, twisted like pretzels, or so expensive to build that the power plant would never be finished. It was like designing a house with a perfect living room, only to realize the stairs to get there require a crane the size of a skyscraper.

The New Way: The "Simultaneous Dance"

Recently, engineers tried doing both steps at once (Single-Stage Optimization). They tried to design the bubble and the magnets together.
The Problem: This is like trying to solve a Rubik's Cube while juggling chainsaws. It's incredibly hard to calculate, often gets stuck in "dead ends," and most existing tools can only handle simple wire coils, not the new, fancy types of magnets (like permanent magnets) that companies are now using.

The Solution: The "Quasi-Single-Stage" (QSS) Proxy

This paper introduces a clever middle-ground called QUADCOIL. Think of it as a flexible, smart "shadow" or "proxy" for the magnets.

Instead of trying to build the actual 3D magnets during the design phase (which is slow and messy), the computer creates a smooth, continuous "ghost sheet" of current around the plasma.

• The Analogy: Imagine you are trying to design a path for a river. Instead of digging the actual riverbed (which is hard), you draw a smooth line on a map that represents where the water should flow. You optimize the shape of the river based on this smooth line. Once you have the perfect river shape, you then go back and figure out how to build the actual rocky banks (the real magnets) to match that line.

Why is QUADCOIL Special?

1. It's Flexible: It can model different types of magnets, including the new "Permanent Magnet" arrays (like thousands of tiny fridge magnets) that companies like MUSE are using.
2. It's Fast & Smart: It uses a mathematical trick called "differentiation" (like a super-smart GPS) to instantly tell the designer: "If you tweak the shape of the bubble just a tiny bit, the magnets will become 10% easier to build."
3. It's Differentiable: This is a fancy math word meaning the computer can calculate the "slope" of the problem instantly. It doesn't have to guess and check; it knows exactly which direction to move to make the magnets simpler.

The Results: Real-World Wins

The authors tested this "ghost sheet" method on two real projects:

• Project 1: The MUSE Experiment (Permanent Magnets)

  • Goal: Reduce the number of giant magnets needed.
  • Result: By using the QUADCOIL proxy, they found a new shape for the magnetic bubble that required 29% fewer magnets than previous designs. That's like saving millions of dollars and tons of steel just by changing the shape of the invisible bubble.

• Project 2: The ARIES-CS Reactor (Filament Coils)

  • Goal: Reduce the massive physical force (stress) the magnets feel. If the magnets push too hard against each other, they could break.
  • Result: The new design reduced the peak force on the magnets by 31%. This makes the reactor much safer and cheaper to build because the magnets don't need to be reinforced as heavily.

The Big Picture

This paper is like giving architects a magic blueprint tool. Instead of building the whole house and then realizing the stairs are too steep, the tool lets them tweak the floor plan while they are drawing the stairs, ensuring the final building is both beautiful (good physics) and buildable (simple, cheap magnets).

It bridges the gap between "theoretical perfection" and "engineering reality," making the dream of a fusion power plant much closer to becoming a reality.

Technical Summary

1. Problem Statement

Designing a stellarator power plant requires balancing plasma performance (physics) with coil engineering costs (complexity, forces, and manufacturability).

• The Two-Stage Limitation: Traditional designs use a two-stage approach: Stage 1 optimizes the plasma equilibrium, and Stage 2 designs coils to reproduce that field. This often results in equilibria with excellent physics properties but coils that are too complex, expensive, or physically impossible to build.
• The Single-Stage Challenge: Single-stage optimization (simultaneously optimizing plasma and coils) offers a better balance but faces significant hurdles:
  • Most tools are specialized only for filament coils, failing to model Permanent Magnets (PM) or dipole arrays (crucial for modern compact designs like MUSE).
  • They suffer from numerical instability, non-convexity, and high computational costs.
  • They often require maintaining thousands of degrees of freedom (coil geometry + plasma shape), making convergence difficult.
• The Gap: There is a lack of a flexible, differentiable "proxy" that can link the plasma equilibrium stage to realistic coil constraints (like force limits or magnet density) without the full computational burden of a single-stage filament optimization.

2. Methodology: Quasi-Single-Stage (QSS) with QUADCOIL

The authors propose a Quasi-Single-Stage (QSS) optimization framework that integrates a coil complexity subproblem directly into the equilibrium optimization loop. The core innovation is the development of QUADCOIL, a differentiable coil proxy based on the winding surface model.

Key Components:

• Winding Surface Model: Instead of modeling discrete wires, the coil set is treated as a continuous current sheet ($K'$) on a "winding surface" surrounding the plasma. This transforms the coil design problem into a linear least-squares problem (or a Quadratically Constrained Quadratic Program - QCQP), which is much faster and often convex.
• QUADCOIL Code: The authors reformulated the winding surface model into a QCQP framework. Unlike previous methods (e.g., REGCOIL) that only minimize current norms, QUADCOIL supports:
  • Non-convex quadratic objectives: Such as minimizing peak Lorentz forces or dipole density.
  • Hard constraints: Allowing users to specify exact targets for metrics (e.g., maximum force, minimum coil spacing).
  • Differentiability: The entire pipeline is differentiable using JAX (auto-differentiation) and adjoint methods, allowing gradients of the coil proxy to be backpropagated to the plasma parameters.
• Differentiable Winding Surface Generator: A new algorithm (Alg. 1) was developed to generate smooth winding surfaces that move with the plasma surface during optimization. It handles self-intersections and preserves "bean-shaped" cross-sections (common in stellarators) while remaining compatible with JAX autodifferentiation.
• Adjoint Differentiation: The paper utilizes the Cauchy Implicit Function Theorem to differentiate the solution of the QCQP subproblem. This allows the calculation of $\nabla f_c(x)$ (the gradient of the coil cost with respect to plasma shape) efficiently, avoiding the need for finite differences.

3. Key Contributions

1. QUADCOIL as a Proxy: Development of a fast, flexible, and differentiable coil complexity proxy that supports realistic engineering constraints (forces, density, topology) unavailable in previous winding surface models.
2. Differentiable Optimization Pipeline: Implementation of an end-to-end differentiable QSS framework using JAX, enabling gradient-based optimization of plasma shapes based on complex coil metrics.
3. Robust Surface Generation: A new differentiable algorithm for generating winding surfaces that avoids self-intersections and preserves geometric features, overcoming limitations of previous offset methods.
4. Superior Efficiency: Demonstrated that QSS with QUADCOIL converges significantly faster and with fewer resources than existing single-stage or parameter-scan-based QSS methods (e.g., MUSE++).

4. Numerical Results

The authors validated the proxy through two distinct case studies:

Study 1: Permanent Magnet (PM) Optimization for MUSE

• Goal: Reduce the number and density of permanent magnets required for the MUSE stellarator.
• Method: Optimized the plasma equilibrium to minimize dipole density ($\Phi'$) and count while maintaining field accuracy.
• Results:
  • The QSS-optimized equilibrium required 34% fewer magnets than the previous MUSE++ design to achieve the same field accuracy.
  • The QSS equilibrium achieved a 47% reduction in peak dipole density and a 61% reduction in total dipole count compared to MUSE++.
  • Efficiency: The QSS optimization took 10–40 minutes on a single consumer GPU (RTX 4060), compared to 12 hours on 16 CPUs for the MUSE++ approach.

Study 2: Low-Force Filament Coils for ARIES-CS

• Goal: Reduce the Lorentz forces on filament coils for the reactor-scale ARIES-CS stellarator.
• Method: Used a QUADCOIL proxy minimizing the $L_1$ norm of the Lorentz force subject to field error and current density constraints.
• Results:
  • The new equilibrium achieved a 27.6% reduction in RMS coil force and a 31.3% reduction in peak coil force compared to the original ARIES-CS design.
  • Plasma physics properties (rotational transform, quasi-symmetry) remained nearly identical.
  • Efficiency: Converged in ~4.7 hours on a single GPU, significantly less resource-intensive than typical single-stage optimizations.

5. Significance and Conclusion

• Bridging the Gap: This work successfully bridges the gap between high-fidelity plasma physics and practical coil engineering. It proves that optimizing the plasma shape with coil constraints in mind (via a proxy) yields significantly better engineering outcomes than optimizing physics first and coils second.
• Versatility: The framework is not limited to filament coils; it successfully handles Permanent Magnets and dipole arrays, making it applicable to next-generation compact stellarators.
• Computational Efficiency: By using a winding surface proxy within a QSS loop, the method achieves single-stage benefits (better coil-plasma balance) at a fraction of the computational cost, running on consumer hardware rather than supercomputers.
• Future Outlook: The authors suggest that QSS solutions can serve as excellent initial states for full single-stage optimizations, potentially accelerating convergence and further improving the design of fusion power plants.

In summary, the paper introduces QUADCOIL, a powerful tool that enables the simultaneous optimization of stellarator plasma and coil systems, leading to designs that are physically robust and significantly cheaper to engineer.