Deflation Techniques for Stellarator Equilibrium and Optimization

This paper introduces a deflation technique that modifies the objective function to penalize previously found solutions, enabling the efficient discovery of multiple distinct, high-quality stellarator equilibria and coil designs from a single initial guess without relying on prescient starting points.

The Gist

Imagine you are trying to design the perfect shape for a futuristic fusion reactor, a machine that aims to replicate the power of the sun. This machine, called a Stellarator, is incredibly complex. It's like trying to sculpt a piece of clay that is constantly trying to reshape itself, all while you are blindfolded and only allowed to touch it in a few specific spots.

The goal is to find the "perfect" shape that traps hot plasma (the fuel) without letting it leak out or damaging the walls. However, the landscape of possible shapes is like a vast, foggy mountain range filled with thousands of valleys. Some valleys are deep and beautiful (great designs), but many are shallow, dead-end pits (bad designs).

The problem is that when you try to find the best shape using computers, the software usually gets stuck in the first valley it finds. It doesn't matter how hard you push it or how you tweak the starting point; it tends to climb out of one small hole only to fall right back into the same one, or a very similar one.

The Problem: Getting Stuck in the Same Hole

Traditionally, to find a different valley, scientists had to guess wildly different starting points or change the rules of the game (the "weights" of their objectives) thousands of times. This is like trying to find a new valley by randomly jumping off a cliff in different directions. It's expensive, time-consuming, and often you just end up landing in the same spot again.

The Solution: The "Magic Eraser" (Deflation)

This paper introduces a clever new trick called Deflation.

Think of the computer's search process like a hiker looking for the lowest point in a valley.

1. The First Hike: The hiker starts, finds a great valley, and says, "Okay, this is a good spot."
2. The Problem: If the hiker tries to start again from the same spot, they will just walk right back into that same valley.
3. The Deflation Trick: Now, imagine the hiker places a giant, invisible inflatable airbag (or a force field) right on top of the valley they just found.
   • If the hiker tries to walk back toward that old valley, the airbag pushes them away.
   • This forces the hiker to look elsewhere.
   • Because the airbag is there, the hiker is now forced to explore a completely different part of the mountain range.
   • Once they find a new valley, they place another airbag on top of it.

By stacking these "airbags" (mathematically called deflation operators) on top of every solution they've already found, the computer is physically prevented from returning to old answers. It is forced to keep exploring until it finds a brand new, distinct valley.

What They Discovered

Using this "Magic Eraser" technique, the researchers were able to:

• Find Hidden Families of Designs: They discovered that for a single set of requirements, there isn't just one "best" shape. There are entire families of different shapes that work just as well. It's like realizing there are many different ways to build a perfect house, not just one blueprint.
• Discover "Helical Cores" Without Guessing: In the past, to find a specific type of twisted, helical shape inside the reactor, scientists had to manually twist the starting shape with their hands (a "prescient guess"). With deflation, the computer found these twisted shapes on its own, simply because the "airbag" pushed it away from the standard, straight shapes.
• Optimize Coils Efficiently: The outside of a Stellarator has complex magnetic coils (like giant electromagnets). Finding the right shape for these coils is a nightmare. Deflation helped them find six completely different, high-quality coil designs from a single starting point, giving engineers more options to choose from.

Why This Matters

This method is like giving the computer a "second chance" to look at the problem without forgetting what it already learned. It's easy to use (you just add a rule to the existing software) and it's very powerful.

Instead of blindly throwing darts at a board hoping to hit a new spot, deflation lets the computer systematically clear away the spots it's already hit, ensuring it explores the whole board to find the absolute best, most diverse set of solutions. This brings us one step closer to building a clean, limitless energy source.

Technical Summary

Here is a detailed technical summary of the paper "Deflation Techniques for Stellarator Equilibrium and Optimization" by Panici et al.

1. Problem Statement

Stellarator optimization and non-axisymmetric equilibrium solving are inherently multi-objective, non-convex problems characterized by complex objective landscapes containing numerous local minima.

• Sensitivity: The solution obtained from a single optimization run is highly sensitive to the initial guess, objective function weights, and the optimization algorithm used.
• Limitations of Current Methods: Conventional approaches to finding distinct solutions involve running optimizations with varied initial conditions or hyperparameters. However, these methods often fail to converge to physically distinct minima, frequently returning to the same local minimum or failing to converge entirely, even after large-scale parameter scans.
• Specific Challenges:
  • Equilibrium Solving: Multiple valid equilibrium states can exist for the same fixed-boundary inputs (e.g., axisymmetric vs. helical core states), but solvers typically find only one based on initialization.
  • Optimization: Finding multiple high-quality local minima in stage-one (equilibrium shape) and stage-two (coil design) optimizations is computationally expensive when relying on random restarts.

2. Methodology

The paper introduces and applies deflation techniques to stellarator physics. Deflation modifies the problem formulation to "penalize" or "deflate" regions around already-found solutions, forcing the solver to seek new, distinct minima.

Core Deflation Mechanism

For a system of equations $F(x) = 0$ or an optimization problem $\min g(x)$, once a solution $x^*_i$ is found, the system is modified using a deflation operator $M(x; x^*_i)$:
$$M(x; x^*_i) = \left( \frac{1}{\|x - x^*_i\|_2^p} + \sigma \right)$$
Where $p$ is a power parameter and $\sigma$ is a shift parameter.

• For Equation Solving: The system becomes $M(x; x^*_i)F(x) = 0$. This creates a barrier around the known root, preventing the solver from converging to it again from the same initial guess.
• For Optimization: The deflation operator is formulated as a nonlinear inequality constraint ($M(x; x^*_i) \leq r$). This forces the optimizer to stay outside a specific region surrounding known solutions.

Key Methodological Innovations

1. Iterative Deflation: The process is repeated iteratively. After finding a solution $x^*_1$, the system is deflated to find $x^*_2$, and so on.
2. State Selection: The deflation operator can be applied to the full state vector or specific subsets (e.g., only the magnetic axis geometry or boundary Fourier coefficients) to target specific physical distinctions.
3. Sum-Reduction vs. Product-Reduction:
   • Standard deflation multiplies operators for multiple solutions (Product-Reduction). The authors identified a flaw where adding more solutions could inadvertently shrink the excluded region around previous solutions due to the multiplicative nature of the constraint.
   • Novel Contribution: The authors propose a "Sum-Reduction" deflation operator ($\sum M(x; x^*_i)$) for optimization constraints. This ensures that the exclusion zones around known solutions remain distinct and do not interfere with each other as the number of deflated solutions grows.

3. Key Contributions

• First Application to Stellarators: While deflation has been used in tokamak Grad-Shafranov solvers, this is the first application to stellarator equilibrium and optimization.
• Discovery of Equilibrium Families: Demonstrated the ability to find families of global equilibria with similar core characteristics but distinct outer geometries and shear profiles.
• Automatic Helical Core Discovery: Successfully found helical core equilibria (non-axisymmetric states with fixed axisymmetric boundaries) without requiring prescient, manually perturbed initial guesses. The deflation process itself generated the necessary perturbation.
• Robust Multi-Solution Optimization: Showed that deflation can efficiently generate multiple high-quality, distinct solutions for both stage-one (quasi-helical symmetry) and stage-two (coil) optimization problems from a single initial guess.
• Algorithmic Improvement: Introduced the "sum-reduction" operator to solve the issue of shrinking exclusion zones in multi-solution optimization.

4. Results

The methodology was tested using the DESC (Direct Equilibrium Solver Code) stellarator code.

A. NAE-Constrained Equilibrium (Near-Axis Expansion)

• Setup: Solved for 25 distinct quasi-axisymmetric (QA) equilibria constrained by Near-Axis Expansion (NAE) theory.
• Outcome: Found 25 distinct solutions sharing the same magnetic axis and on-axis rotational transform but exhibiting different rotational transform profiles, shear signs, and outer flux surface geometries.
• Physics: Some solutions possessed stable magnetic wells, while others did not, providing a diverse set of initial guesses for further optimization.

B. Helical Core Equilibria

• Setup: Applied to a fixed-axisymmetric boundary problem known to support both axisymmetric and helical core states.
• Outcome: Starting from a standard axisymmetric guess, the deflated solve identified a perturbed initial condition that converged to the helical core equilibrium. This was achieved without manually designing the perturbation, proving deflation's utility in bifurcation detection.

C. Stage-One Optimization (Quasi-Helical Symmetry)

• Setup: Optimized a vacuum stellarator for quasi-helical symmetry (QH) with constraints on aspect ratio and rotational transform.
• Outcome: Found 18 distinct equilibria passing all constraints. While some were physically similar due to Fourier parametrization degeneracy (toroidal shifts), the method successfully explored a wide range of rotational transform profiles and boundary shapes.

D. Stage-Two Optimization (Filamentary Coils)

• Setup: Optimized coil sets for a precise QA stellarator using the sum-reduction deflation operator.
• Outcome: Successfully generated 6 distinct coil sets.
  • 5 of the 6 sets satisfied all geometric and field error constraints.
  • The sets exhibited different normal field errors and coil geometries.
  • One set (Index 4) was visually distinct but failed constraints, highlighting the method's ability to explore the full solution space, including infeasible regions.

5. Significance and Future Work

• Efficiency: Deflation offers a computationally cheaper alternative to running thousands of random restarts to explore the optimization landscape. It allows for the systematic discovery of "families" of solutions.
• Ease of Implementation: The method requires minimal changes to existing codes (DESC), typically just adding a constraint or modifying the objective function.
• Future Directions:
  • Improved Formulations: Adopting least-squares specific deflation algorithms (e.g., from Riley et al.) to improve convergence robustness.
  • Parametrization: Addressing Fourier representation degeneracy (e.g., using exponential spectral scaling or reduced state variables like magnetic well depth) to ensure mathematically distinct solutions are also physically distinct.
  • Hyperparameter Tuning: Systematic scanning of deflation parameters ($p, \sigma, r$) to control the diversity of the solution space.

In conclusion, this paper establishes deflation as a powerful, versatile tool for navigating the complex, non-convex landscapes of stellarator physics, enabling the discovery of diverse equilibrium states and optimization solutions that were previously difficult to access.