Sophelio has launched the Fusion Equilibrium Challenge, an open scientific machine-learning competition accepted for the NeurIPS 2026 Competition Track.
The challenge invites machine-learning researchers, data scientists, physicists, engineers, and students worldwide to address a problem central to the future of fusion energy:
Can machine learning estimate the geometry and key equilibrium properties of a fusion plasma without using magnetic diagnostics as model inputs?
No previous fusion background is required, and all challenge data, tools, baselines, and documentation are available openly.
A reactor-relevant machine-learning problem
Today’s tokamaks commonly rely on magnetic measurements as inputs to equilibrium-reconstruction systems. In future reactor environments, however, intense neutron exposure may degrade magnetic sensors or make them more difficult to maintain and rely upon.
The Fusion Equilibrium Challenge asks whether important plasma-equilibrium information can instead be inferred from diagnostic and control signals expected to remain available in reactor-class devices.
Participants receive poloidal-field coil currents, plasma current, Thomson-scattering electron-temperature and density profiles, and fixed machine geometry. Their models must predict a full (65×65) poloidal-flux map, (ψ(R,Z)), together with q95 and normalized beta. Plasma-boundary and additional geometric quantities are then derived from the predicted flux map during evaluation.
Real experimental data from two tokamaks
The open dataset contains 9,121 experimental plasma shots and 98 GB of data from two substantially different operating tokamaks:
- DIII-D, operated by General Atomics in the United States
- MAST, operated by the United Kingdom Atomic Energy Authority (UKAEA)
These are real experimental measurements—not simulations. The dataset is available on Hugging Face under a CC BY 4.0 license.
Two challenges
The competition includes two separately awarded tracks.
Challenge 1: DIII-D intra-machine inference
Develop and evaluate an equilibrium-inference model on data from the same tokamak.
Challenge 2: DIII-D-to-MAST cross-machine generalization
Train using DIII-D data and estimate MAST equilibria without using MAST training data.
The cross-machine challenge tests whether a model has learned transferable equilibrium structure rather than patterns specific to a single experimental device. This is the more difficult challenge—and potentially the more consequential one for future fusion systems.
Free access to dFL
To make the challenge accessible to the widest possible community, Sophelio is providing dFL, its multimodal scientific data-labeling and processing platform, free of charge for challenge participants.
dFL allows participants to open, inspect, label, compare, and process the heterogeneous diagnostic signals contained in the challenge dataset without having to build a complete data-exploration environment from scratch. It is designed to help both fusion experts and newcomers understand the data, investigate individual plasma shots, and prepare information for machine-learning workflows.
Together with dFL, participants receive four reference baselines, sample data, complete documentation, and a step-by-step getting-started guide designed to take users from initial setup to a first submission quickly.
Participation is open
Phase 1 is open now and closes October 18, 2026.
A blind final phase will run from October 19 through October 26, with results presented through the NeurIPS 2026 Competition Track in December.
Each challenge carries a $500 award.
The Fusion Equilibrium Challenge was developed with contributions from Sophelio, DIII-D/General Atomics, the United Kingdom Atomic Energy Authority, and the University of Texas at Austin.
- Explore the challenge, access the data and dFL, and submit a model:
- Visit fusion-equilibrium-challenge.sophelio.io
A fuller announcement describing the scientific motivation, collaborators, supporting organizations, and long-term goals of the challenge will follow.