Skip to content
SIH Buddyby Ganeev Singh
Dev
All problem statements
SIH26078Proceed with cautionacceptance 2/5

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

Ministry of Earth Sciences (MoES) · Disaster Management · Software

A genuinely well-argued research pipeline that is simply too large — the diffusion downscaler needs compute and data you will not have, so if you take this, scope openly to the tracking stage rather than promising the half you cannot deliver.

What it actually is

Finding where an extreme weather event will actually hit inside a huge global forecast dataset is still largely done by eye, and the coarse grid these models run on smears out exactly the sharp peaks of rain and wind that matter. The ask is a system that automatically finds and tracks the anomaly across a three-to-ten-day window, then sharpens it to a finer grid without flattening the extremes.

What to build

The two-stage pipeline the statement lays out: a graph neural network operating on an icosahedral mesh rather than a flat pixel grid, computing an extreme forecast index against a long reanalysis climatology to isolate anomalies and draw a moving bounding box around each one across the forecast window; and a conditional generative diffusion model taking that cropped region and downscaling it from the coarse ensemble grid to a finer one, trained to generate physically plausible high-amplitude fields rather than to minimise mean error, since minimising mean error is precisely what blurs the peaks — with physics-informed penalties in the loss for states that violate conservation, plus a visualisation layer and an alerting interface that drops a coordinate at the anomaly centre with graded severity.

Smallest thing that wins the room

Take a documented historical cyclone, show the tracker's bounding box following it across the forecast window, then put your downscaled field beside a conventional regression downscaling of the same slice and show the peak intensities surviving in yours and disappearing in theirs.

How crowded this one gets

A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.

Moderate140–330 teams expectedroughly 1 in 119–275 wins it

Quieter than 42% of the 226 · #131 of 226 by expected field

A normal-sized field. Your idea has to be good, not miraculous.

Why: central ministry statements sat below the average.

This is a guess, not a fact

Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.

The scores

The number is the shorthand. The line under it is the reason.

What you will be writing

  • message-passing GNN on icosahedral mesh
  • conditional denoising diffusion downscaling
  • Extreme Forecast Index against ERA5 climatology
  • physics-informed conservation loss terms
  • Xarray and Dask on 4D NetCDF/GRIB2 arrays
  • amplitude-preserving verification against regression baseline
  • Numerical weather prediction post-processing
  • Generative modelling for geosciences
  • Extreme event tracking

Prior art to read before you start

statistical downscaling preserving extreme amplitudes · graph neural networks on spherical grids · ensemble anomaly detection and tracking

Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.