Bringing Fusion Down to Earth: ML for Stellarator Optimization
At Hugging Face, we believe that solving the world’s hardest scientific problems shouldn’t be limited to closed labs or siloed institutions. What if we could open up fusion research—the quest for clean, limitless energy—to the entire machine learning community? We’re excited to announce a new collaboration with Proxima Fusion, a spin-out from the Max Planck Institute for Plasma Physics and the fastest-growing fusion company in Europe. Together, we’re launching a series of open challenges to accelerate fusion engineering with machine learning—starting with a simulation-driven approach to stellarator optimization.
Because the future isn’t just imagined–it’s modeled, optimized, and built.
Part 1: The Context
What’s fusion?
For decades, controlled nuclear fusion has been a beacon promising clean, safe, and virtually limitless energy. Fusion powers the stars by fusing light nuclei into heavier ones, releasing enormous energy. But confining hot plasma—a searing soup of charged particles—is unsurprisingly nontrivial. The two leading contenders for a magnetic confinement approach to fusion are tokamaks and stellarators. While tokamaks have long led the race, stellarators have re-emerged as a promising alternative due to recent advances in design and simulation, novel experimental results, and powerful magnets.
What’s a stellarator?
A stellarator is a type of fusion device designed to confine hot plasma using twisted, three-dimensional magnetic fields to sustain nuclear fusion reactions.
Unlike tokamaks, which use a combination of external magnets and a current driven through the plasma, stellarators rely entirely on external magnets to shape and confine the plasma. This makes them inherently more stable and suitable for continuous (steady-state) operation, since they avoid issues like current-driven instabilities and disruptions. However, the magnetic field geometry in stellarators is much more complex, requiring advanced optimization and coil design to achieve good confinement.
A Terminology Primer for the ML Community
Quasi-isodynamic (QI)
Quasi-isodynamic (QI) fields are a class of magnetic configurations in stellarator design that improve particle confinement by ensuring that charged particles remain close to their original magnetic flux surfaces, even when they are trapped. Unlike in conventional stellarators where trapped particles tend to drift away and cause energy losses, QI configurations reduce this drift by approximating a conserved quantity called the longitudinal adiabatic invariant. This leads to better confinement performance without requiring the symmetry of a tokamak, making QI a key property for next-generation stellarators.
Wendelstein 7-X and the Stellarator Breakthrough
The Wendelstein 7-X (W7-X) experiment is the world’s most advanced stellarator, designed and built by the Max Planck Institute for Plasma Physics (IPP) in Greifswald, Germany. Its construction began in 2005, and the device was officially completed in 2014, with first plasma achieved in December 2015.
W7-X was conceived as a proof of concept: could a stellarator be designed and built to confine plasma as effectively as a tokamak, but without the instability and pulsed operation challenges that plague tokamaks? Its mission was to demonstrate that steady-state magnetic confinement could be achieved in a carefully optimized 3D magnetic field geometry, leveraging decades of theoretical and modeling progress.
To build W7-X, researchers used a computational optimization process to design the magnetic field coils that shape the plasma. The resulting device features a complex, twisted geometry that can maintain stable confinement without relying on externally-induced currents in the plasma. It was, in many ways, a gamble on whether high-performance fusion could be achieved through geometrical optimization.
The payoff came in 2018 and 2022, when experiments showed that W7-X could achieve record-low levels of neoclassical transport—meaning the plasma stayed confined far better than any previous stellarator. 2022 was truly an “annus mirabilis” for stellarators, with newly published theoretical results suggesting that a massive acceleration could take place—if we could simulate and design stellarators fast enough.
But that “if” is precisely the challenge. The design and simulation pipeline behind W7-X involved massive computational effort, many design iterations, and hand-tuned design parameters. To build next-generation stellarators faster and better, we’ll need new approaches—and that’s where the ML community might be able to help.
What's slowing us down?
Stellarators are notoriously difficult to build, largely due to their intricate 3D geometry. Traditional computational pipelines rely on physics solvers like VMEC and HINT, which compute the equilibrium between the magnetic field, but these solvers can be slow and fragile. Moreover, highly optimized devices like W7-X often require millimeter-level tolerances, which can derail cost and schedules due to manufacturing complexity. Could optimized QI stellarators help overcome these engineering challenges? As the class of configurations with the lowest physics risk on the path to grid-ready fusion, QI stellarators offer a promising foundation. If we can significantly simplify their engineering design, we can accelerate the path to practical fusion energy.
Part 2: The Challenge
We propose three stellarator optimization problems of increasing complexity, each with progressive relevance to fusion reactor design:
Geometrically Optimized Stellarator – Design a shape that minimizes elongation under fixed constraints on aspect ratio, triangularity, and magnetic field twist.
Simple-to-Build Quasi-Isodynamic (QI) Stellarator – Optimize for plasma shapes that promote good confinement via the QI property, but are simpler to build, with smoother magnetic fields requiring simpler coils.
Multi-Objective, MHD-Stable QI Stellarator – Map the compactness and simplicity tradeoff while ensuring confinement and stability.
Each benchmark problem comes with reference implementations, evaluation scripts, and strong baselines in using classical optimization methods using the Proxima open source codebase.
What does the data look like?
The ConStellaration dataset contains over 150,000 QI equilibria produced by VMEC++. As a reminder, QI stellarators are a subset of configurations that minimize the internal plasma currents that can lead to disruptive events in a tokamak. The provided equilibria correspond to different 3D plasma boundary surfaces and offer samples across a wide and physically meaningful range of parameters. The dataset includes:
- Input parameters: the 3D plasma boundary, together with the pressure and current profiles.
- Equilibrium outputs: full VMEC++ equilibrium solution plus additional metrics of interest for stellarator design (e.g., degree of QI symmetry, turbulent transport geometrical quantities).
What are we designing?
This challenge asks participants to optimize stellarator designs using ML, e.g., by building surrogate models that can predict the outcome of VMEC++ simulations and key downstream qualities from input parameters. These models could ultimately replace expensive simulations in the stellarator design pipeline, enabling real-time design iteration and differentiable optimization loops.
We are hosting a live leaderboard, where researchers can submit optimized designs and compare performance on standard evaluation metrics.
What would it unlock?
Fusion offers a zero-carbon, fuel-abundant, and intrinsically safe energy source that could transform our global energy system. Unlike fossil fuels, it produces no greenhouse gases. Unlike fission, it produces no long-lived radioactive waste. And unlike solar and wind, it’s not intermittent.
But we won’t get there without a new generation of tools—ones that let us simulate, optimize, and design fusion reactors orders of magnitude faster. Combining physics with machine learning can accelerate this timeline.
Part 3: Call for Contribution
We’re inviting the machine learning and fusion communities to join forces. This challenge is just the beginning. We’re looking for contributions across the stack:
- Submit boundaries to the leaderboard to test and benchmark your approach to optimizing the plasma.
- Build new architectures or surrogate modeling techniques tailored to physical simulation tasks.
- Contribute to the dataset by sharing new VMEC++ simulations, augmenting the coverage of equilibrium regimes or extending the parameter space.
- Collaborate with us–we’re actively looking for partners in academia, industry, and open science who want to accelerate progress in fusion engineering through ML.
Whether you're an ML researcher looking for high-impact scientific applications, a plasma physicist curious about differentiable optimization, or a student exploring fusion for the first time—there’s a place for you in this collaboration. We’ll provide support, tools, documentation, and open discussions to help contributors onboard and stay engaged. If you're interested in contributing, reach out to us or clone the Constellaration dataset on Hugging Face and get started.
Conclusion: Why It Matters
Controlled fusion has long been the holy grail of energy research. It's safe. It’s abundant. And it could power the world for millions of years without polluting it. But unlocking it requires not just better physics—but also better tools. This collaboration is one small step toward that goal: turning fusion design into a fast, iterative, and ML-native process. By making simulation data public, defining benchmark tasks, and inviting the ML community in, we hope to accelerate the timeline for practical fusion energy. We believe that with the right tools—and the right people—what was once a decades-away dream could become a reality much faster than most people think.
Get Started
from datasets import load_dataset
# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("proxima-fusion/constellaration", "default")
More Resources
To get a better understanding of the state of the field and the problem at hand, we recommend the following resources:
