Germany's Proxima Fusion Releases New Benchmark for Stellarator Coil Design
Recently, Proxima Fusion, a German stellarator device development company, officially released an open dataset and machine learning benchmark named CoilStellaration. This achievement successfully transforms the complex underdetermined inverse problem of deriving coils from target magnetic fields for stellarators into an open, comparable data-driven generation task, marking a key step forward for stellarators toward commercializable, engineerable design.

For stellarators to move toward commercial fusion, in addition to solving the plasma equilibrium optimization problem, the practical coil design challenge of generating confining magnetic fields must be overcome. Since deriving coils from target magnetic fields is a classic underdetermined inverse problem, infinitely many coil geometries can produce the same magnetic field. To this end, CoilStellaration released by Proxima Fusion contains 177,341 optimized stellarator coil sets (totaling 266,549 samples), covering representative conditions from small experimental devices to reactor-scale machines, and forms a seamless connection with the previously jointly developed ConStellaration, comprehensively bridging the complete chain from plasma shape to practical coil engineering implementation.
In terms of technical implementation and evaluation framework, the dataset is generated through a two-stage process of “REGCOIL fast warm start + DESC formal optimization,” while balancing stringent engineering constraints such as magnetic field accuracy, total coil length, coil spacing, coil-to-plasma distance, and maximum curvature. Its core benchmark task requires machine learning models to predict complete coil sets in a single forward pass, without relying on expensive optimization loops during inference. Currently published benchmark tests indicate that the design is quite challenging and far from saturated, providing a unified performance evaluation arena for researchers worldwide.
Industry experts note that the establishment of this open benchmark has profound industry significance: it not only compresses traditional days-long coil design exploration to hours, providing a high-quality “warm start” starting point for subsequent optimization, but also greatly lowers the barrier for the machine learning community to participate in fusion engineering design. By deeply integrating cutting-edge AI technology with open-source stellarator tool ecosystems such as DESC and SIMSOPT, CoilStellaration is building solid design infrastructure for globally buildable, economically viable stellarator fusion power plants.
Disclaimer: Information republished from partner media, institutions or other websites is provided for reference and communication purposes only. It does not imply endorsement of its views or verification of its accuracy. Please contact us if any content infringes rights or requires correction.