01
Active monsoon
Widespread, sustained rainfall across the subcontinent. Strong moisture flux along the monsoon trough.
Broad correction
SMART INDIA HACKATHON 2026
SIH26080
Smarter monsoon forecasts.
One regime at a time.
ATMOS · Adaptive Targeted Monsoon Optimization System
ATMOS post-processes existing NWP rainfall forecasts using regime-aware AI — identifying whether conditions are active, a break, a depression, or in transition, then applying the right correction. It does not replace the physical forecast model.
04Regime experts
05Pipeline stages
01Clearer picture
CONCEPT VISUALIZATION — NOT A LIVE FORECAST
02 / THE CHALLENGE
Forecast errors change
with the weather.
Numerical Weather Prediction (NWP) models produce rainfall forecasts for the entire Indian subcontinent. But a single global bias-correction cannot account for how error patterns shift across different monsoon conditions.
During an active monsoon surge, errors may be driven by widespread convection. During a break, rainfall retreats to the foothills. A depression or low-pressure system brings organised, intense rainfall in a compact area. And transition periods are uncertain by nature — no single pattern dominates.
Correcting all four situations the same way likely misses the structure of each error. ATMOS proposes to recognise the regime first, then apply the appropriate correction — or blend multiple corrections when the situation is genuinely uncertain.
PROBLEM STATEMENT
MoES / NCMRWF · SIH26080
A single correction model is applied to all situations. It may work well on average but miss regime-specific error patterns. This is a conceptual illustration — not a performance claim.
ATMOS estimates regime probabilities and blends specialist corrections accordingly. Uncertainty is kept visible rather than hidden. This is a conceptual illustration — not a performance claim.
02
Break period
Rainfall retreats to the Himalayan foothills and northeast. Central India experiences a dry spell.
Targeted correction
03
Depression / Low
Organised low-pressure system produces intense, spatially complex rainfall in a compact region.
Concentrated correction
04
Transition
The monsoon is changing state. No single regime dominates — uncertainty favours blended experts.
Blended experts
03 / OUR PROPOSED SOLUTION
From raw NWP
to district intelligence.
ATMOS post-processes the physical forecast — it does not replace it. Select any stage to understand its purpose.
CENTRAL INNOVATION
Soft gating blends four regime specialists and a global fallback, weighted by estimated regime probabilities and uncertainty — so the system adapts to every atmospheric situation rather than forcing a single approach.
PIPELINE / 01
Ingest & align
NWP rainfall forecasts, atmospheric fields (winds, humidity, temperature, vorticity), and terrain data are brought onto consistent spatial grids, accumulation windows, and forecast lead-times. Missing or invalid inputs are flagged before any modelling begins.
Scientific safeguard
Forecast-time features must never include target-day observed rainfall. Verification uses held-out data only.
BUILT IN THIS SHOWCASE
- Vanilla HTML / CSS / JS (no build step)
- Three.js — interactive 3D globe
- CSS animations & scroll reveals
- Deployed on Vercel (static)
PLANNED FOR WORKING PROTOTYPE
- Frontend — React / Next.js · TypeScript · Tailwind
- Maps — Leaflet / MapLibre · GeoJSON district layers
- Backend — Python · Django/DRF platform · FastAPI model serving
- ML — scikit-learn · LightGBM (XGBoost / PyTorch where justified)
- Geo & weather — xarray · NumPy · pandas · GeoPandas · rasterio
- Data formats — NetCDF · GRIB · GeoTIFF
- Storage — PostgreSQL / PostGIS · object storage
- Jobs & deploy — Docker · background task workers
04 / REGIME EXPLORER
Choose a regime.
See how it changes everything.
Each monsoon regime produces a different error pattern in raw NWP. Select a regime to understand its character and why it calls for a specialist approach. Expert weights shown are illustrative examples — not trained-model output.
Active monsoon surge
Strong, widespread rainfall driven by a deep monsoon trough. The southwest flow carries abundant moisture from the Arabian Sea. Broad convective activity.
Typical NWP error tendency
NWP models may underestimate peak convective rainfall in heavy-rain events while overestimating lighter organised rain. Spatial displacement errors can occur near the Western Ghats orographic zones.
ILLUSTRATIVE EXPERT WEIGHTS
These percentages are illustrative examples only. Actual weights will come from a trained soft-gating model.
05 / PRODUCT VISION
The working prototype
will look like this.
An interactive decision-support workspace — compare raw NWP with corrected rainfall, inspect heavy-rain probability and uncertainty, and navigate from grid to district.
Illustrative interface preview
Real model and forecast data will be connected in the working prototype. No district value, rainfall amount, probability, or accuracy figure shown here is measured or official.
This interface preview is illustrative. Geography and visual layers are conceptual. No forecast date, measured value, or official alert is implied. A real URL will be linked here when the working prototype launches.
06 / FEASIBILITY & VIABILITY
Built on proven
foundations.
ATMOS proposes to combine well-established NWP, machine-learning, and GIS technology. Each element rests on publicly available methods and candidate data sources.
NWP availability
ECMWF open-data and IMD-NCMRWF products are candidate NWP sources. Alignment of forecast grids, lead-times, and accumulation windows is a tractable data engineering problem.
Observation access
IMD gridded rainfall data and station observations are candidate verification and training targets. Legal access terms, temporal alignment, and spatial coverage must be confirmed for any training set.
ML feasibility
Gradient-boosted trees and ensemble blending are well-tested for post-processing tasks. The regime classification step draws on established monsoon indices and circulation diagnostics from the literature.
Geographic refinement
High-resolution terrain data (SRTM / ASTER DEM) and coastal distance fields are publicly available. GIS tooling for orographic parameterisation is well established.
Key challenges
NWP model changes, non-stationary monsoon climatology, extreme-event rarity, and data quality all require careful handling. Scaling to more districts, leads, and thresholds depends on data availability.
Integration path
Post-processing can run as a downstream step once a NWP forecast is issued. No modification to the physical model is required. Output can feed district-level decision-support interfaces.
07 / INTENDED IMPACT
Better rainfall intelligence
for better decisions.
Outcomes are intended benefits pending validation against held-out observations. No measured improvement is claimed.
01
Disaster preparedness
More spatially accurate district rainfall forecasts could support earlier mobilisation for floods and landslides — pending validation of forecast improvement.
02
Agriculture & water planning
District-level rainfall summaries and percentile information could inform irrigation scheduling and reservoir management decisions.
03
Localised forecast information
Area-weighted grid-to-district aggregation with p90/p95 and fraction-above-threshold provides richer information than point or mean forecasts.
04
Uncertainty-aware prediction
Probabilistic heavy-rain estimates and visible regime uncertainty help decision-makers understand confidence levels rather than receiving a single deterministic number.
05
Forecast quality assessment
The built-in verification framework enables systematic comparison between raw NWP, global correction, and ATMOS using standard meteorological skill scores.
06
Decision support tool
Risk indicators are decision support for informed human judgment — not official IMD warnings. The system is designed to complement, not replace, operational forecasting services.
All impact claims are anticipated benefits from a proposed system.
No bar chart, accuracy percentage, or measured improvement from this site
should be interpreted as a validated experimental result.
08 / SCIENTIFIC INTEGRITY
Improvement is
a measurement.
The working prototype will compare three systems on held-out observations. Until that comparison is done, no improvement claim belongs on this page.
VERIFICATION STATUS
Results pending model training and validation
KEY SAFEGUARD
Forecast-time features must never include target-day observed rainfall. Verification must use held-out data that was not seen during training. Risk indicators are decision support and not official IMD warnings.
01
Raw NWP
The unmodified physical model forecast — the baseline to beat.
02
Global correction
A conventional, regime-agnostic bias-correction model — the fair comparison point.
03
ATMOS
Regime-aware blended correction — the proposed improvement to evaluate.
CONTINUOUS
RMSE · MAE · Bias
How close are predicted rainfall amounts to observations?
EVENT
POD · FAR · CSI · ETS
How well are heavy-rainfall threshold events identified?
SPATIAL
FSS
Does the corrected forecast capture the observed rainfall pattern at useful spatial scales?
PROBABILITY
Brier Score · Reliability
Do estimated heavy-rain probabilities match observed exceedance frequencies?
Results pending measured validation
Numerical skill scores will appear in the working prototype after the data pipeline is aligned, models are trained, and held-out verification is complete.
09 / THE PEOPLE
Meet Team
MEDWRAP.
Six people connecting weather research, data engineering, AI, geospatial analysis, and a clear experience for decision-makers.
TEAM LEAD
GUIDANCE / FACULTY MENTORS
Guided by
experience.
Technical guidance and review from our faculty mentors.
11 / RESOURCES & CONTACT
More to come.
Stay curious.
Explore the project video, documentation, and prototype website. Research references below are attributed candidate sources.
Contact the team leader
WhatsApp · Pre-filled ATMOS enquiry
Heavy rain coast to coast — I'm working overtime! The active expert takes the lead here.