Diakhère Gueye
← All work

Designing an international ML challenge at the IAEA

How do you take a hard, expert-only research question in fusion science and open it up so that hundreds of data scientists around the world can enter, and learn from it?

My role
Challenge design, data harmonisation, documentation & public delivery
With
IAEA · ITU AI for Good · MIT PSFC · HUST · SWIP
Reach
452 participants · three tokamaks · two continents

The problem

Tokamaks are the machines fusion research uses to hold plasma hot enough to fuse. Sometimes that plasma disrupts: it goes unstable, dumps its energy, and the run ends, potentially damaging the machine. Predicting a disruption a moment before it happens is one of the field's open problems. The catch is that every tokamak is different, so a model trained on one tends to fail on another. The genuinely useful, and genuinely hard, question is cross-machine generalisation: can a model learn disruption behaviour that transfers to a machine it has never seen?

The IAEA wanted to open that question to the wider machine-learning community, not just fusion physicists. My job was to make that possible.

What I did

I worked within a team spanning the IAEA, ITU, MIT's Plasma Science & Fusion Center and two Chinese fusion institutes, and I owned the part that turns a research idea into a challenge people can take part in.

  1. Framed the task

    Around cross-machine generalisation rather than single-machine prediction, because that is where the real difficulty and the real value sit, and it is a question an outsider can grasp without a physics degree.

  2. Made three very different machines comparable

    Harmonising experimental data from J-TEXT and HL-2A in China and Alcator C-Mod at MIT into one common format.

  3. Designed the evaluation

    Participants trained on two machines and were scored on the held-out third, so the leaderboard measured transfer to an unseen device, not memorisation.

  4. Wrote the documentation

    So a data scientist with no fusion background could understand the data, the task and the rules well enough to compete.

  5. Brought people in

    Ran it publicly and spoke on the ITU AI for Good webinars that introduced the challenge and its results to an international audience.

J-TEXT training machine HL-2A training machine model Alcator C-Mod held-out test, never seen
The evaluation in one picture: train on two machines, score on the third the model has never seen, so the leaderboard measured genuine transfer, not memorisation.
The achievement here wasn't a model. It was lowering the barrier: shaping the framing, the data, the evaluation and the documentation so that hundreds of people who had never touched a tokamak could make a genuine contribution.

Outcome

The challenge drew 452 participants and produced a public benchmark for cross-machine disruption prediction, backed by five research institutions across two continents. You can still see it on Zindi and through ITU AI for Good (I'm listed on the ITU speaker page).

Most of the work in getting a technical system adopted lies outside the system itself: the framing, the fairness of the evaluation, and whether a newcomer can get started at all. Building a model is the straightforward part. The genuinely hard, decisive work is everything around it, and designing that is exactly what I do best.

← Back to all work