CTF4Nuclear is a NeurIPS 2026 competition for evaluating machine-learning surrogates of nuclear-reactor multiphysics simulations. The first challenge in the series scores models on the Molten Salt Fast Reactor (MSFR) — a Generation-IV reactor concept whose liquid-fuel design couples neutronics, thermal-hydraulics, and salt transport in ways that make full-physics simulations expensive. Models are evaluated on twelve standardised experiments covering short- and long-time forecasting, denoising, limited-data, and parametric generalisation; the public [Hugging Face leaderboard](https://huggingface.co/spaces/ctf4science/ctf4nuclear-msfr-leaderboard) runs continuously on Hugging Face compute.
The challenge
Nuclear fission and fusion reactors are among the most safety-critical dynamical systems ever engineered. Their safe operation demands continuous, real-time knowledge of the full multi-physics state of the reactor core — neutron flux distributions, fuel and coolant temperatures, coolant velocity fields, and delayed-neutron precursor concentrations. In next-generation reactor concepts — Generation-IV fission designs such as molten salt reactors and micro-reactors, and magnetic-confinement fusion devices — extreme environments make in-core instrumentation nearly impossible: sensors are physically confined to core boundaries or external shielding structures.
Competition task. Reconstruct the spatially resolved, multi-field state of the reactor core given only sparse point measurements of a single observable field. This setting is fundamentally different from standard time-series forecasting and has no analogue in existing ML benchmarks. The first challenge in the CTF4Nuclear series targets the Molten Salt Fast Reactor: five coupled fields — prompt fission power, decay heat, temperature, and two velocity components — on a 3 880-node mesh, evaluated across twelve experiments that probe short- and long-time forecasting, denoising, limited-data, and parametric generalisation.
Why it matters. Digital twins capable of real-time state estimation from sparse sensors are a prerequisite for autonomous reactor control, predictive maintenance, and rapid accident prognosis. The IAEA has identified AI-enabled digital twins as a strategic priority for nuclear safety, and the OECD Nuclear Energy Agency has called for standardised ML benchmarks in this domain. The underlying inverse problem — recovering a high-dimensional multi-physics state from partial, indirect observations — also arises in climate modelling, industrial process control, and biomedical monitoring, making algorithmic advances broadly transferable.
Tentative timeline
The launch date and the NeurIPS Competition Workshop date are fixed. All other dates below are tentative and will be finalised at competition launch.
| Date | Milestone |
|---|---|
| June 2026 | Competition opens — public Hugging Face leaderboard accepts submissions |
| August 2026 | Mid-competition mini-talks (top participants present their approaches) |
| Mid-October 2026 | Public Hugging Face leaderboard closes; final phase begins (top submissions re-evaluated on organiser hardware against a held-out parametric regime) |
| November 2026 | Winners announced; PMLR proceedings write-ups invited |
| 11–12 December 2026 | NeurIPS Competition Workshop, Sydney |
How to participate
- Create a Hugging Face account (if you don’t have one).
- Clone the
ctf4sciencepackage andpip install -e .. - Download the training data from
ctf4science/ctf4nuclear-msfr. - Generate a 9 500-row predictions Parquet (or run the starter notebook to get a working Baseline Last submission).
- Push the Parquet to your own public Hugging Face dataset repository.
- Submit via the Submit tab on the Hugging Face leaderboard — scores appear within minutes.
The starter notebook completes steps 2–6 end-to-end in under ten minutes on a free Colab CPU instance.
Rules
The competition rules are tentative and still being finalised. Winning entries are released under an OSI-approved open-source license, and winning teams are invited to contribute a short write-up to the PMLR competition-track proceedings volume.
Full official rules — including submission caps, allowed external data and tools, and the final re-evaluation procedure — will be published on the Hugging Face leaderboard before competition launch.
Frequently asked questions
Who can participate? Anyone with a Hugging Face account.
Do I need a GPU? Not necessarily. Submissions are prediction files, not training code, and the Hugging Face leaderboard scoring runs entirely on Hugging Face compute — no resources are required from you to submit. However, most deep-learning approaches will still need a GPU to train your model. The reference Baseline Last is pure NumPy and runs in under ten minutes on a free Colab CPU instance.
Where is the training data? Public mirror at
huggingface.co/datasets/ctf4science/ctf4nuclear-msfr.
Backup tarball at OSF.
What format is the submission? A single Parquet file with 9 500 rows
and columns id, pair_id, timestep, f1, …, f19400. Schema and
per-pair row layout: see the Submission Format tab on the
Hugging Face leaderboard.
How is my score computed? Twelve metrics across forecasting, denoising, and parametric generalisation, each clipped to [−100, 100], averaged into AvgScore. Formal definitions: ctf4science documentation.
Can I submit multiple times? Yes — the per-day and final-submission limits will be set in the official rules published before launch.
Is there a private leaderboard? Yes. After the public phase closes, top finishers are re-evaluated on organiser hardware against a held-out parametric regime; the private scores are the official final rankings.
How do I cite this benchmark? A reference paper “CTF4Nuclear: A Common Task Framework for Nuclear Fission and Fusion Models” is under review at NeurIPS 2026; a citation will be added here on publication.
Where can I ask questions? Use the Community tab on the Hugging Face leaderboard. For email, write to alexeyy@uw.edu for technical questions (submission, scoring, data) or to kutz@uw.edu for general questions about the competition.
Lead organisation and sponsors
The CTF4Nuclear competition is led by the AI Institute in Dynamic Systems, with co-organisers across the University of Washington, Columbia University, Politecnico di Milano, MIT, American University of Beirut, University of Cambridge, Autodesk Research, Distyl AI, and SURF; the full organising team is listed in the companion paper. Hugging Face sponsors all submission-evaluation compute and hosts the public Hugging Face leaderboard.