AI-Driven Spatio-Temporal Tracking of Extreme Weather Anomalies in Medium-Range Forecasts

The receipts 🧾

Every red flag on this statement, in full. These are the four places it bites.

1. Exhibit A

   Training a diffusion model on atmospheric fields is a research-lab undertaking in both compute and data volume, and this is the stage that actually solves the stated problem — a submission with a working tracker and a conceptual downscaler has answered half the statement

2. It gets worse

   Historical global ensemble archives from the sponsoring centre are not openly distributed, so the forecast inputs the pipeline is designed around may simply be unavailable to you

3. Still reading?

   Innovation scope is effectively zero because the architecture is fully written, so every responsive team builds the same pipeline and only execution differs

4. And the finisher

   Generative downscaling produces plausible high-amplitude fields, which is not the same as correct ones — be clear that the diffusion stage generates a physically consistent realisation rather than a deterministic prediction, because conflating those is the error a forecaster will catch

The damage report 📉

Every score this statement earned, and what each one actually costs you.

• Feasibility

  2/5

  You have picked a fight with physics, procurement, or both. One of them always wins.

  This is a research-laboratory pipeline written as a problem statement — training a conditional diffusion model on four-dimensional atmospheric fields against decades of reanalysis is a serious GPU budget and a multi-terabyte data problem, and access to historical global ensemble archives from the sponsoring centre is not something a student team can arrange in the time available.

• Innovation scope

  1/5

  Nothing here is new. Your only edge is execution — and execution is also everyone else's only edge.

  Every design decision is already made in the statement — the icosahedral mesh, the message-passing graph network, the extreme forecast index against a reanalysis baseline, the conditional denoising diffusion downscaler, the physics-informed loss penalties and even the software libraries are all specified.

• Clarity

  5/5

  The ask is unambiguous, which quietly removes your favourite excuse.

  Exceptionally detailed, correctly identifying the spectral smoothing problem that motivates the whole approach, specifying both architecture stages, naming the frameworks, the baseline and forecast datasets, the resolutions and all four deliverables.

• Acceptance potential

  2/5

  The numbers do not like you. Bring something the numbers cannot see.

  The compute and data volumes put this beyond what a student team can execute, the architecture is fully dictated so there is nothing to invent, and most submissions will implement the tracking stage adequately while the generative downscaler — the part that solves the stated problem — remains a slide.

• Effort

  Massive

  A semester of work wearing a hackathon costume. Something is getting cut; decide what now, not in week five.

  A graph network on a spherical mesh, a conditional diffusion model with custom physics-guided losses, a multi-terabyte four-dimensional data pipeline, a visualisation layer and an alerting API is a research programme, and either stage alone would be a full project.

• Demo-ability

  Medium

  Demoable, if you rehearse it. Nobody rehearses it.

  The side-by-side showing preserved peak amplitudes against a blurred conventional downscaling is a genuinely good comparative visual and makes the whole argument in one frame, but it requires the diffusion stage to actually work, which is the least likely part to.

• Data

  None supplied

  No dataset comes with this one, so every accuracy figure you quote is a number about labels you invented.

  Nothing is provided with the statement. You are sourcing, cleaning and labelling it yourself, and that work is invisible in the demo but very visible in the questions.

The demo they will have already seen 🪞

Somewhere around 140–330 teams are heading here, and the description is doing the choosing for most of them. They will read the same brief, reach the same architecture, and build a version of the same demo you are planning. Being correct is the floor. If your five minutes could be swapped with the team before you and nobody in the room would notice, you have not picked badly — you have built predictably, which costs exactly the same and hurts more.

What survives 🛡️

The ground worth standing on when the questions start.

• The statement correctly identifies a real and well-known failure of conventional downscaling — that optimising mean error blurs exactly the peaks forecasters need — and that framing gives you a sharp, defensible argument and an obvious comparative demo
• The tracking stage alone is genuinely achievable on a reduced domain and is a complete, useful contribution if you scope to it honestly
• Reanalysis climatology for computing an extreme forecast index is freely available, so the anomaly detection half has no data barrier