NeurIPS 2026 — Real Government Fusion Reactor Data · $500 Awards · Your Name on the Paper

This is a real scientific competition, accepted to NeurIPS 2026 — the Conference on Neural Information Processing Systems, the world's most cited machine learning conference, where foundational breakthroughs in deep learning, generative AI, and scientific ML have been presented for over three decades. Being part of its competition track puts your work in front of the global ML research community. Two $500 awards. Your name on a peer-reviewed paper. An invitation to present at NeurIPS 2026 in December. The problem Future fusion reactors — SPARC, ARC, ITER — will operate under neutron fluxes that degrade their magnetic sensors. Without those sensors, traditional plasma control systems lose their primary inputs. The Fusion Equilibrium Challenge asks whether machine learning can fill that gap: reconstruct the full magnetic geometry of a fusion plasma using only the diagnostics future reactors will still have. This is not a toy dataset. This is the actual scientific challenge facing the next generation of energy infrastructure. Built with - General Atomics — operator of DIII-D, the largest magnetic fusion research facility in the US, funded by the US Department of Energy, home to 800+ scientists from 100+ institutions worldwide - UKAEA — the UK government's national fusion research authority, data released through its FAIR-MAST open data program - University of Texas at Austin — Institute for Fusion Studies, research co-organizer - Hugging Face — the world's leading open-source AI platform, hosting the dataset - NeurIPS — the world's most cited ML conference Your submission is evaluated against researchers and engineers worldwide. Your name goes on the paper if you place. The data 9,121 real experimental plasma shots from two government-operated fusion facilities — DIII-D (General Atomics, San Diego) and MAST (UKAEA, Culham, UK). 98 GB. Open on Hugging Face under CC BY 4.0. Not simulations — real operational fusion diagnostics collected over decades of publicly funded research. What you build A harmonization layer + an ML model predicting the full 2D plasma flux map and key equilibrium parameters. What you win $500 — best DIII-D reconstruction $500 — best zero-shot transfer to MAST Named co-authorship on the lessons-learned paper Invited talk at NeurIPS 2026 Skills that win Neural operators · Transfer learning · Physics-informed ML · PyTorch / TensorFlow · Python No fusion background required. Four reference baselines included. First submission in under an hour. Start here - Download dFL free — visualize every plasma shot before you write a line of training code: 🔗 dfl.sophelio.io - Full challenge details, dataset, rules, and official announcement: 🔗 fusion-equilibrium-challenge.sophelio.io Phase 1 open through October 18, 2026.