Artificial intelligence company Sophelio has launched the Fusion Equilibrium Challenge, inviting machine learning researchers to develop models capable of reconstructing the magnetic structures that confine plasma inside fusion reactors without relying on magnetic sensor data.
The competition has been accepted to the NeurIPS 2026 Competition Track and is being organised in partnership with the DIII-D National Fusion Facility, UK Atomic Energy Authority’s FAIR-MAST programme and the University of Texas at Austin’s Institute for Fusion Studies.
Sophelio describes it as the first fusion energy competition accepted to the NeurIPS Competition Track.
At the centre of the challenge is a problem facing the development of future fusion systems. Tokamaks use powerful magnetic fields to confine extremely hot plasma, and understanding the geometry of those fields is important for monitoring and controlling the plasma.
Traditional equilibrium reconstruction methods depend on magnetic measurements. However, the intense neutron environments expected inside future reactor-class systems can degrade magnetic sensors, potentially making conventional approaches more difficult to maintain over extended operations.
The competition asks researchers whether AI can reconstruct this magnetic geometry using non-magnetic information instead.
Participants will use measurements including poloidal-field coil currents and Thomson-scattering electron profiles to predict a complete two-dimensional poloidal flux map and key equilibrium parameters. The objective is to determine whether machine learning can infer the magnetic structure even when direct magnetic measurements are unavailable or degraded.
Sophelio has released 133 GB of experimental tokamak data for the competition. The dataset contains 9,121 curated plasma discharges from two fusion research machines, including 7,915 from the DIII-D National Fusion Facility in San Diego and 1,206 from MAST through UKAEA’s FAIR-MAST programme.
The inclusion of two machines also creates a second test for participating AI models. Researchers will examine whether models trained using DIII-D data can generalise to MAST without additional training, despite differences in machine geometry, diagnostics and operating conditions.
The competition therefore consists of two main challenges. The first evaluates reconstruction performance within DIII-D, while the second measures zero-shot transfer from DIII-D to MAST.
The dataset has been released under a Creative Commons CC BY 4.0 licence and is hosted on Hugging Face. Participants can also use Sophelio’s Data Fusion Labeler to inspect plasma data, including flux contours, electron profiles and diagnostic time series.
The development phase runs until October 18, with a public leaderboard allowing participants to evaluate their models. A blind final phase is scheduled from October 19 to October 26, followed by verification and peer review.
Final results are expected to be presented during NeurIPS 2026 in December. Two cash awards of $500 each will recognise the strongest DIII-D reconstruction and cross-machine generalisation results. Leading teams can also receive co-authorship on a planned lessons-learned paper and invitations to present their work.
The competition reflects the growing application of machine learning to scientific research beyond conventional generative AI use cases. In fusion research, AI is increasingly being investigated for areas including plasma modelling, prediction and control.
Whether models developed through the competition can eventually support reactor-scale operations remains to be established. For now, the benchmark provides researchers with an open dataset for testing whether AI can infer critical plasma information when conventional magnetic measurements are limited.