🔬 physics

Towards Stellarator Geometry Optimisation for Nuclear Fusion

This paper presents a local geometry refinement method and a 2D latent representation that achieved top performance on the geometric task of the ConStellaration leaderboard in May 2026 for stellarator geometry optimization.

Original authors: Tobias Weißberg, Moritz Heep, Zorah Lähner, Florian Bernard

Published 2026-08-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tobias Weißberg, Moritz Heep, Zorah Lähner, Florian Bernard

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

For decades, scientists have chased the dream of fusion power, a clean energy source that mimics the process fueling the sun. To make this work on Earth, researchers must trap superheated gas, known as plasma, which is ten times hotter than the core of our star. One way to hold this searing material without it touching the walls of a container is to suspend it within a twisted magnetic field. A device called a stellarator uses this method, shaping the magnetic field itself to keep the plasma confined and stable, rather than driving an electric current through the gas as other designs do. The success of a stellarator depends entirely on the precise shape of the magnetic cage; if the geometry is even slightly off, the plasma can escape or become unstable. Designing this shape is a massive challenge because the physics involved are so complex that they cannot be solved with simple formulas, requiring instead powerful computer simulations to test every possible configuration.

In a recent effort to improve these designs, a team of researchers from the University of Bonn and the Lamarr Institute in Germany focused on refining the geometry of stellarator plasma boundaries. They entered a competition known as the ConStellaration benchmark, which challenges scientists to find the best possible shape for a specific type of stellarator. The goal was to minimize how stretched the plasma cross-section becomes while ensuring the overall structure remains compact and properly twisted. The researchers approached this not by starting from scratch, but by taking the best existing solutions from a public leaderboard and applying a new method to polish them. They replaced the standard evolutionary search algorithms, which mimic natural selection to find solutions, with a technique that uses local gradient-based optimization. In plain terms, this method calculates the direction of improvement from a starting point and takes a direct step toward a better shape, repeating this process until no further gains can be found. This approach allowed them to refine five of the top existing designs, improving each one significantly in just a few hours of computation.

Beyond simply polishing existing designs, the team created a new way to visualize and explore the space of possible shapes. They mapped the complex, high-dimensional data of the plasma boundaries onto a simple two-dimensional plane, preserving the relationships between different solutions. This map revealed that the best-performing shapes were not isolated islands but were connected by a smooth landscape of possibilities. By sampling this landscape, the researchers discovered a new, blended geometry that combined features of the top entries in a way that the original optimization methods had missed. This new shape was feasible, meaning it met all the physical requirements, and it scored higher than any previously known solution. When they applied their refinement method to this newly discovered shape, the score improved even further, surpassing all other entries on the leaderboard at the time of their submission.

The work demonstrates that even in a field driven by complex physics simulations, simple geometric refinements and better ways of visualizing the problem space can yield significant improvements. The researchers noted that their method is local, meaning it improves upon a starting point but cannot jump across vast gaps to find entirely new types of solutions from scratch. However, by combining their local refinement with a low-dimensional map of the solution space, they were able to find a superior design that had remained hidden. Their final score of 0.9763 set a new standard for this specific geometric task, proving that careful analysis of the relationships between known solutions can lead to breakthroughs in the design of future fusion reactors.

Technical Summary: Towards Stellarator Geometry Optimisation for Nuclear Fusion

Problem Statement
The paper addresses the geometric optimisation of stellarator plasma boundaries for nuclear fusion. Unlike tokamaks, stellarators rely entirely on the shape of the magnetic field for plasma confinement, making the design of the plasma boundary a critical shape-optimisation problem. The specific challenge tackled is the "geometric task" from the ConStellaration benchmark, a dataset of high-fidelity simulations for Quasi-isodynamic (QI) stellarators. The objective is to minimise the plasma's cross-sectional elongation while satisfying four constraints: maintaining a compact aspect ratio, ensuring appropriate D-shape (average triangularity $\le -0.5$ ), and achieving a specific rotational transform (magnetic field line twist). While the first three metrics depend solely on the boundary surface geometry, the rotational transform requires solving the complex, non-linear Magnetohydrodynamic (MHD) equilibrium, for which no closed-form solution exists.

Methodology
The authors propose a two-pronged approach combining local refinement and global exploration within a learned latent space:

Local Gradient-Based Refinement:
- The authors treat optimisation as a refinement problem for existing leaderboard solutions.
- They replace the standard evolutionary algorithm (CMA-ES) used in the benchmark with a local gradient-based method (L-BFGS).
- Gradients are computed in parallel across a 306-dimensional input space using a forward, two-point finite-difference stencil directly through the VMEC++ solver. This setup utilizes 96 concurrent solver instances per gradient calculation, taking approximately one hour per run.
- The minimisation objective is affinely transformed into a score to be maximised, consistent with the benchmark protocol.
2D Latent Space Representation:
- To explore the solution space beyond local minima, the authors construct a 2D latent representation of the geometry.
- Embedding: They use Multidimensional Scaling (MDS) to embed existing solutions into a 2D plane, preserving pairwise Euclidean distances in the high-dimensional Fourier coefficient space.
- Interpolation: To map points in this 2D plane back to Fourier coefficients, they employ mean value coordinates. The ten embedded solutions form a closed polygon; any point in the plane is mapped to a weighted sum of the Fourier coefficients of the polygon's vertices. This ensures exact reproduction at vertex positions and smooth interpolation elsewhere.
- This representation traces a curved 2D slice through the coefficient space rather than a flat plane, allowing for the sampling of "blended" geometries that are non-linear combinations of existing solutions.

Key Results

Refinement Performance: The local refinement method was applied to the top-5 public entries on the ConStellaration leaderboard (as of May 2026). The method improved all five geometries. Four of the five were improved by at least 115% of the distance between the previous best and the previous 5th solution.
Latent Space Discovery: By sampling the 2D latent space generated by the top-5 entries and their refinements, the authors identified a novel "blended" geometry. This point, representing a recombination of existing solutions, achieved a score of 0.9752, surpassing all known solutions at the time.
Final Score: Further local refinement of this blended design yielded a final score of 0.9763. The authors submitted an entry scoring 0.9743, which topped the leaderboard at the time of submission.
Landscape Analysis: Visualisation of the latent space revealed that the feasible region is connected but contains "infeasible pocket intrusions" (likely due to the discontinuous triangularity metric). The objective function varies smoothly, and the optimal solutions form a band extending beyond the initial refined entries, suggesting that traversing infeasible regions via the latent map is a viable strategy for finding better geometries.

Significance and Claims
The paper claims to present a method that successfully topped the ConStellaration leaderboard for the geometric task with a submitted entry. The significance lies in demonstrating that:

Local refinement is effective: Replacing evolutionary algorithms with parallelised L-BFGS on the specific benchmark task yields substantial improvements over existing solutions.
Latent space sampling adds value: A low-dimensional, non-linear representation based on MDS and mean value coordinates allows for the discovery of superior geometries through recombination (blending) that local line-search methods cannot propose.

Limitations and Scope
The authors maintain a modest scope regarding their claims:

Local Nature: The refinement method is inherently local; it inherits the basin of attraction of the starting solution and requires a high-quality initial solution to refine. It did not reach the best-known solution (0.9763) independently without the latent space sampling, nor was this best-known solution submitted to the leaderboard.
Task Scope: The work focuses exclusively on the geometric task. The other two benchmark tasks (involving more complex physics) are acknowledged as natural continuations but were not addressed.
Computational Cost: The method relies on costly finite-difference evaluations. The authors note that integrating the low-dimensional representation directly into the optimisation loop and finding alternatives to finite differences are necessary future steps.

Technical Summary: Towards Stellarator Geometry Optimisation for Nuclear Fusion

More like this