A supercomputer simulation that predicts future weather by solving fluid-dynamics equations on a global 3D grid. ECMWF and GFS are the leaders; Mausam Online uses ECMWF via Open-Meteo.
forecastingWhat is a Weather Model?
A numerical weather prediction (NWP) model — informally called a “weather model” — is a computer program that simulates the atmosphere to predict future weather. Modern models divide the atmosphere into a 3D grid of millions of cells and solve the laws of physics (mainly the Navier-Stokes equations of fluid dynamics, plus thermodynamics, radiation transfer, moisture, and surface interactions) for each cell at each timestep.
The atmospheric simulation is initialized with current observations from:
- Weather stations (~10,000 worldwide)
- Radiosondes (weather balloons released twice daily from ~800 sites)
- Aircraft (AMDAR sends in-flight observations — tens of thousands daily)
- Satellites (geostationary + polar-orbiting; the dominant data source)
- Buoys, ships, drifters over oceans
- Radar for precipitation and wind
The model then integrates forward in time, producing forecast output every 1-6 hours out to 10-14 days. Beyond that, individual weather becomes unpredictable (the “butterfly effect”); only general atmospheric trends can be forecast at longer ranges.
The major global weather models
ECMWF IFS (European Centre for Medium-Range Weather Forecasts) — IFS = Integrated Forecasting System:
- Resolution: 9 km horizontal, 137 vertical levels, 1-hour timestep
- Forecast range: 10 days deterministic, 46 days ensemble
- Generally considered the most accurate global model
- Used by Mausam Online via Open-Meteo
- Headquartered in Reading, UK with supercomputer in Bologna, Italy
- Members include India, Sri Lanka, all EU states
GFS (US Global Forecast System):
- Resolution: 13 km horizontal, 127 levels, 4-times-daily cycle
- Forecast range: 16 days deterministic
- Operated by NOAA/NCEP
- Free public data — used by many weather apps
- Slightly less accurate than ECMWF on average
ICON (German Weather Service):
- 13 km global, 6.5 km Europe nest
- Strong over central Europe
- Public data available
UKMet (UK Met Office Unified Model):
- 10 km global
- Particularly skilled at North Atlantic weather
- Not publicly available (commercial)
JMA (Japan Meteorological Agency):
- 13 km global, 5 km regional
- Excellent for typhoon forecasting
IMD GFS (Indian Meteorological Department):
- Customized GFS implementation for South Asia
- 12.5 km horizontal
- Used by IMD for official forecasts
- Includes WRF (Weather Research and Forecasting) regional models
HARMONIE/AROME (European regional model):
- 2.5 km horizontal — high resolution
- Used by national met services in Nordic countries, Netherlands, France
Regional and high-resolution models
Global models lack the resolution to capture local-scale weather like thunderstorms, mountain rain or sea breezes. Limited-area models (LAMs) nest inside global models with higher resolution:
- WRF (Weather Research and Forecasting) — used by IMD, PMD, BMD at 4-12 km
- HRRR (US High-Resolution Rapid Refresh) — 3 km, updated hourly
- ICON-D2 (German national) — 2.2 km
- AROME-France — 1.3 km
For Indian monsoon forecasting, the NCMRWF (National Centre for Medium Range Weather Forecasting) operates an Indian-tuned GFS at 12 km plus WRF nested at 4 km for severe events.
Ensemble forecasting
A single model run is deterministic — one specific forecast. But the atmosphere is chaotic; small initial errors grow over time. Ensemble forecasting runs the same model 30-50 times with slightly perturbed initial conditions and physics parameters. The spread of outcomes tells us how confident the forecast is.
- ECMWF ENS: 51 members
- GFS GEFS: 31 members
- High agreement = high confidence forecast
- Wide spread = low confidence; multiple scenarios possible
For monsoon, cyclones and heatwaves, ensembles are essential. ECMWF’s seasonal forecasts (S2S) extend ensemble forecasting to weeks/months.
How Mausam Online uses NWP
Mausam Online’s forecasts come from Open-Meteo, a free weather API that aggregates output from multiple top-tier models:
- Primary: ECMWF IFS (9 km)
- Backup: GFS, ICON
- Hourly resolution
- 7-day forecast
- Updates every 1-3 hours
Open-Meteo specifically uses the ECMWF AIFS (AI-enhanced) and HRES (high-resolution deterministic) outputs, with downscaling for South Asian terrain. This makes the forecasts you see on Mausam Online directly comparable to those used by professional meteorologists.
Accuracy and limits
Today’s NWP skill (approximate):
- Day 1: 95%+ accurate for temperature, 80%+ for precipitation
- Day 3: 85% temperature, 65% precipitation
- Day 5: 75% temperature, 50% precipitation
- Day 7: 60% temperature, 35% precipitation
- Beyond 10 days: only general trends
What models struggle with:
- Convective precipitation — individual thunderstorms hard to place precisely
- Mountain rainfall — terrain effects at fine scale
- Coastal sea breeze — small-scale circulations
- Tropical cyclone landfall point — track errors of 50-100 km even at 1-3 day lead
- Snow forecasting — temperature thresholds critical and small errors matter
Climate change and NWP
Climate change is reshaping NWP in several ways:
- More frequent extreme events strain models trained on historical patterns
- AI-based models (ECMWF AIFS, Google GraphCast, Huawei Pangu-Weather) emerging as competitors to physics-based NWP
- 2024-25 AIFS results show AI models can match or exceed physics models on some metrics
- Coupled ocean-atmosphere-land models needed for accurate seasonal/decadal projections
- Higher resolution (1-2 km globally) anticipated within a decade
Frequently asked questions
Why are weather forecasts wrong sometimes? The atmosphere is chaotic — small initial errors grow exponentially over time. By 7-10 days, individual weather features are essentially unpredictable. Model resolution also limits detail at small scales. And convective phenomena (thunderstorms) inherently have lower predictability than synoptic-scale features.
Which weather model is most accurate? ECMWF IFS is consistently ranked highest in formal verification studies. GFS is close behind and free. For specific phenomena (e.g., West Pacific typhoons), JMA model may outperform; for European storms, ICON or UKMet.
Why do different apps give different forecasts? Because they use different models, different data sources, and different interpretation/display logic. Mausam Online uses ECMWF via Open-Meteo. AccuWeather, The Weather Channel, etc. each have proprietary blends.
Can AI replace physics-based models? Increasingly the answer is yes — Google’s GraphCast, Huawei’s Pangu-Weather, and ECMWF’s AIFS demonstrate that AI models trained on past data can match or exceed physics models on many metrics. The future of operational weather forecasting may be hybrid AI/physics.
Where can I see model output for my city? Mausam Online provides ECMWF-based hourly forecasts on every city page. For raw model output, see Windy.com (multiple model selector), ECMWF Open Charts, or NOAA Weather Models page. Live forecasts: Delhi, Mumbai, Chennai, Karachi, Dhaka.