A proof-of-concept for automated AI-driven stellarator coil optimization with in-the-loop finite-element calculations

This paper presents an automated, end-to-end "runner" system for stellarator coil optimization that utilizes genetic algorithms and context-aware LLMs to streamline design workflows, features an open-source leaderboard for tracking solutions, and introduces novel in-the-loop finite-element calculations for Von Mises stress analysis.

The Gist

Imagine you are trying to build a giant, invisible donut out of pure magnetic energy. This donut, called a stellarator, is designed to trap super-hot plasma (like the stuff inside the sun) so we can create clean, limitless fusion energy.

The problem? To hold this magnetic donut together, you need to wrap it in a complex web of superconducting coils (giant metal rings). But these rings can't just be simple circles. They have to twist, turn, and weave in 3D space like a pretzel made of spaghetti to keep the plasma stable.

Designing these coils is like trying to solve a 3D puzzle where the pieces keep changing shape, and if you get it wrong, the whole machine fails. Traditionally, this takes human experts years of work, tweaking numbers by hand, and running massive simulations.

This paper introduces a new "robot architect" that automates this entire process. Here is how it works, broken down into simple concepts:

1. The "Runner": A Self-Driving Car for Design

Think of the old way of designing coils as driving a car where you have to manually adjust the steering, gas, brakes, and mirrors every second while reading a map. It's exhausting and slow.

The authors built a "Runner" (an automated software system). You just tell it: "Build a coil system for a donut-shaped plasma of this size."

• No Manual Tuning: The robot handles the messy setup. It automatically draws the initial coils so they don't crash into the plasma or tangle with each other.
• Auto-Pilot: It runs 24/7, testing thousands of designs without needing a human to press "start" every time.

2. The "Brain": Two Ways to Think

The Runner needs a brain to decide which designs to try next. It uses two different strategies, like a team of two different engineers:

• The Genetic Algorithm (The "Evolutionary Gardener"): Imagine you have a garden. You plant seeds (designs), see which ones grow the best, cut off the weak branches, and cross-pollinate the strongest ones to make new, better seeds. The robot does this mathematically. It takes the best coil designs, mixes them up, adds tiny random changes (mutations), and sees if the new "offspring" is better.
• The LLM (The "Experienced Librarian"): This is a Large Language Model (like a super-smart AI that has read every paper ever written about stellarators). Instead of just guessing, it acts like a seasoned expert. It looks at the history of what worked and what failed, reads the "library" of past physics papers, and says, "Hey, we tried twisting the coils this way before and it broke. Let's try twisting them that way instead." It makes intelligent guesses based on context.

3. The "Safety Check": The Stress Test

This is the paper's biggest breakthrough. Usually, engineers design the coils, and later (months later) they check if the metal will snap under pressure.

The authors added a "Stress Sensor" that works in real-time.

• The Analogy: Imagine you are sculpting a statue out of clay. Usually, you finish the statue and then ask a structural engineer, "Will this fall over?"
• The New Way: As you sculpt, the robot instantly calculates the pressure on every part of the clay. If a part looks like it's about to crack, the robot immediately reshapes that part while it is still designing.
• The Result: They successfully minimized the Von Mises stress (a fancy way of saying "how much the metal is about to snap"). They found designs that were not only good at holding the plasma but were also physically stronger and less likely to break.

4. The "Leaderboard": A Global Competition

To make sure everyone is playing fair, they built an open-source online leaderboard.

• Think of it like a high-score screen in a video game.
• Any researcher in the world can download the code, run their own design, and submit their score.
• Because everyone uses the exact same rules and computer tools, it's a true "apples-to-apples" comparison. If you beat the score, you are genuinely better.

Why Does This Matter?

• Speed: What used to take years of human labor can now happen in days or weeks.
• Reliability: By checking the physical stress during the design, they avoid creating "theoretical" coils that would shatter in real life.
• Democratization: It lowers the barrier to entry. You don't need a PhD in plasma physics to start testing new ideas; you just need to know how to use the leaderboard.

In a nutshell: This paper is about handing the keys of the fusion reactor design to a smart, tireless robot that can design, test, and stress-check the machine's most difficult parts (the coils) all by itself, making the dream of fusion energy a little closer to reality.

Technical Summary

1. Problem Statement

Designing stellarator fusion reactors is a critical challenge for realizing future nuclear power plants. The process involves a vast parameter space and is traditionally divided into two stages:

1. Stage I: Optimizing the plasma boundary shape.
2. Stage II: Optimizing the magnetic coil shapes and currents to reproduce that boundary.

Key Challenges:

• Ill-posed Inverse Problem: Many distinct coil configurations can generate similar magnetic fields, but most are engineeringly impractical (e.g., intersecting coils, excessive curvature, or structural failure).
• Manual Bottlenecks: The workflow requires extensive manual intervention for initialization, weight tuning for multi-objective optimization, and complex post-processing (structural analysis, stress calculations) using disparate simulation codes.
• Computational Cost: Running high-fidelity simulations (MHD equilibrium, Finite Element Method for stress) is expensive, making it difficult to explore the parameter space thoroughly without automation.
• Lack of Standardization: Comparing different coil solutions is difficult due to varying code implementations, plasma surfaces, and resolution parameters.

2. Methodology

The authors developed an end-to-end automated "runner" (code available at stellcoilbench) that integrates initialization, optimization, and post-processing into a continuous loop.

A. Automated Workflow Components

• Automated Initialization: Instead of manual tuning, the system initializes circular coils that are guaranteed to avoid interlinking or intersections with the plasma surface. It automatically tunes total current to match the desired magnetic field.
• Penalty-Based Optimization: The system uses an Augmented Lagrangian method. This eliminates the need for users to manually define objective weights; instead, constraints (e.g., minimum coil-coil distance) are handled via penalty methods that auto-tune.
• Automated Post-Processing: The system automatically runs a suite of well-benchmarked codes (SIMSOPT, VMEC, SIMPLE, ParaStell, scikit-fem/DOLFINx) to validate:
  • Shape gradients and sensitivity analysis.
  • Poincaré plots and Boozer coordinates (for quasi-symmetry).
  • Fast-ion losses.
  • Finite Element Method (FEM) calculations for structural stress and displacement.

B. AI-Driven Policy Algorithms

The system employs a "proposer" to decide the next batch of optimization runs. Two policies are implemented:

1. Deterministic Genetic Algorithm (GA): Uses log-normal mutation and log-uniform exploration to navigate the parameter space (e.g., number of coils, Fourier orders).
2. Context-Aware Large Language Model (LLM): The LLM acts as an "AI expert," accessing a structured knowledge base of previous runs, failure statistics, and domain literature. It reasons about which parameter spaces to explore, adjusts constraint thresholds based on failure modes, and maintains a running log of justifications for reproducibility.

C. In-the-Loop Finite Element Calculations

A novel contribution is the integration of Von Mises stress calculations directly into the optimization loop.

• Implementation: Uses open-source libraries (scikit-fem or DOLFINx) to solve linear elasticity problems on a tetrahedral mesh of the coil winding pack.
• Optimization Tricks: To ensure speed, the system uses low-resolution meshes during the loop, caches stiffness matrices, employs adaptive mesh refinement, and uses "short-circuits" to skip expensive gradient calculations when stress is low.
• Physics: Calculates body forces ($J \times B$) using regularized internal field models to avoid singularities near the coil centerline.

3. Key Contributions

1. Fully Automated Runner: A proof-of-concept system that automates the entire stellarator coil design lifecycle from initialization to leaderboard submission, requiring only basic input parameters from the user.
2. First In-the-Loop Stress Optimization: The first demonstration of directly minimizing Von Mises stresses and coil displacements within the optimization loop using open-source FEM, rather than relying on post-hoc analysis with closed-source commercial software (e.g., ANSYS).
3. AI-Driven Policy: Implementation of a context-aware LLM policy for scientific optimization, capable of learning from failure modes and historical data to propose targeted experiments, alongside a robust Genetic Algorithm.
4. Open-Source Leaderboard: Creation of a standardized, open-source leaderboard where users can submit case.yaml files. This ensures "apples-to-apples" comparisons using identical code, plasma surfaces, and resolution parameters.
5. Code Generation: The majority of the code was generated by LLMs (via the Cursor app), with extensive unit testing and documentation added to verify accuracy.

4. Results

The system was tested on the Landreman-Paul Quasi-Axisymmetric (QA) stellarator (scaled to a 1m major radius).

• Stress Reduction:
  • Optimizing with a penalty on volume-averaged Von Mises stress reduced the peak displacement from ~16 cm (baseline) to ~1 cm.
  • High-resolution post-processing confirmed that in-the-loop optimization resulted in Von Mises stresses an order of magnitude lower than baseline solutions.
• Genetic Algorithm Performance:
  • The unsupervised GA found a highly competitive 3-coil solution (3 coils per half-field period).
  • Compared to a previous 3-coil solution (Gil et al., 2025), the new solution reduced total coil length by ~17% and required ~60 km less superconducting material (75.37 km vs 90.27 km).
  • It achieved a maximum curvature of 0.28 m⁻¹ (vs 0.53 m⁻¹ previously) and doubled the minimum coil-surface distance.
• LLM Performance:
  • The LLM successfully identified under-explored parameter spaces and adjusted constraints dynamically. While the full database generation is ongoing, the system demonstrated the ability to fall back to the GA if the LLM failed, ensuring continuous operation.
• Performance:
  • A fully converged run with default settings took approximately 20 minutes on a Mac M3 MAX using 8 MPI processes.
  • FEM calculations in the loop were accelerated significantly using optimization tricks (e.g., skipping gradients when stress is low).

5. Significance and Future Outlook

• Feasibility of AI in Fusion Design: This work proves that AI agents (both GA and LLM) can autonomously navigate the complex, non-convex landscape of stellarator design, finding solutions that rival or exceed human-engineered designs without supervision.
• Engineering-Physics Integration: By integrating structural FEM directly into the loop, the design process now accounts for mechanical feasibility (stress, displacement) during optimization, not just after. This is a critical step toward designing reactors that are physically buildable.
• Scalability: The framework is modular and can be extended to other codes (e.g., DESC, FOCUS) or other engineering aspects like thermal analysis and neutron flux calculations.
• Reproducibility: The standardized leaderboard and open-source nature of the code address the historical difficulty in comparing stellarator designs, fostering a collaborative environment for rapid iteration.

In conclusion, this paper presents a transformative step toward "Digital Twins" for fusion reactors, demonstrating that automated, AI-driven workflows can significantly accelerate the design of feasible, high-performance stellarator coils while simultaneously optimizing for structural integrity.