KGIndex

India --:--:-- --

All work

03 · Forecast blending · Smart India Hackathon

WEAVR

Weigh every rainfall model by how much it has earned, region by region and lead by lead.

When
2026
Role
ML engineer, team ATMOSYNC
Built with
Python, xarray, Zarr, EMOS, BMA, GraphCast, IFS-ENS
Source
ATMOSYNC/WEAVR ↗

See it

Bar chart of per-region source weights for GraphCast, HRES and the IFS ensemble
Who to trust, region by regionFor each region of India and each lead time, the bars show how much weight the blend gives GraphCast, HRES and the IFS ensemble. Regions without enough training data are marked and fall back to equal weights, and the page says so. Real output of the dashboard on the committed example grids.
Line chart of RMSE by lead time for seven forecasting methods
Every method, scored honestlyRMSE by lead time for every method, from the plain ensemble mean up to regime-conditioned weights. Every method is scored on the same rain-gauge analyses, so the comparison is fair.

The problem

India’s monsoon is forecast by a crowd of models, physics-based and AI. Each is good somewhere, at some lead time, for some kind of rain. Averaging them throws away what is known about each one.

Numbers, with dates

  • 193 → 0false-zero probability cells after the extreme-value tail, at both the 115.6 mm and 204.5 mm thresholdstail-repair-results.md, PR #100
  • 629tests passing, with lint and type checks cleanorg main, 28 Oct 2026
  • 11of my pull requests merged; four more are openGitHub, 29 Oct 2026

How it works

  1. 01 · align

    Put every model on one grid

    Different resolutions, lead times and units become one aligned cube over India.

  2. 02 · weigh

    Learn who to trust where

    Weights are learned per region and per lead from history, not set once for the whole country.

  3. 03 · calibrate

    Turn a number into a probability

    Each calibration method is benchmarked against the last; it only ships if it wins.

  4. 04 · verify

    Verify without leaking

    The year being tested is never seen in training. Every claim in the README links to the run that produced it.

What I built

  • A blend of GraphCast, HRES and the IFS ensemble on a common 0.25° India grid, weighted per region and per lead time.
  • A calibration ladder where each method has to beat the one below it: equal weights, regional weights, EMOS and BMA, then regime-conditioned.
  • Probabilities at the rain thresholds IMD actually warns on, verified against rain-gauge analyses with a leave-one-year-out protocol.

Team

Built as a ATMOSYNC project. The repository credits everyone who worked on it.