Weather prediction has traditionally relied on physical equations solved across a numerical representation of the atmosphere. AI forecasting introduces a complementary approach: learning atmospheric evolution from large historical datasets.

Numerical weather prediction begins with observations, estimates the current state of the atmosphere and advances that state using equations that describe fluid motion, thermodynamics and related processes. It is computationally demanding but grounded in physical theory.

Learning patterns from reanalysis data

AI weather models are commonly trained on reanalysis datasets, which combine historical observations with physical models to create consistent estimates of past atmospheric conditions. The system learns to map one state of the atmosphere to a future state.

Once trained, some AI models can generate forecasts much faster than conventional numerical systems. This creates opportunities for larger ensembles, rapid updates and wider access to forecasting capability.

Forecast skill is not one number

A model may perform well on average variables while missing rare but consequential extremes. Evaluation must consider location, lead time, altitude, season and phenomenon. Tropical cyclones, intense rainfall, heatwaves and local storms present different challenges.

Physical consistency also matters. A forecast can achieve a favourable statistical score while violating relationships that should hold in the atmosphere. Researchers therefore compare AI-based, physics-based and hybrid systems across multiple metrics and events.

Uncertainty is part of the forecast

The atmosphere is chaotic: small uncertainty in initial conditions can grow over time. Operational forecasting uses ensembles to estimate a distribution of possible futures rather than one deterministic answer.

AI systems need credible uncertainty estimates as well. A fast forecast is not sufficient if users cannot distinguish high-confidence conditions from scenarios with substantial spread.

Hybrid systems may combine strengths

AI can accelerate components of physical models, correct systematic errors or produce forecasts that are combined with numerical output. Physics can also constrain learning and help models behave more plausibly outside familiar conditions.

The World Meteorological Organization is developing methods for comparing AI, physics-based and hybrid forecasting systems, including robustness and regional performance. It has emphasised that these technologies should strengthen the complete weather, climate and water service chain rather than remain isolated research demonstrations.

Forecasting for public value

The scientific achievement of faster prediction matters most when it improves warnings and decisions. That requires observations, communication infrastructure, local expertise and institutions capable of translating forecasts into action.

AI is unlikely to make atmospheric physics obsolete. Its more credible role is to become another powerful component of forecasting—tested against physical understanding and judged by whether it improves reliable information for society.

Forecast skill depends on the question

A global average can conceal the event that matters locally. Researchers evaluate forecasts by lead time, variable, altitude, region and event type. Temperature and pressure may be predicted accurately while rainfall extremes or small-scale convection remain difficult. Probabilistic systems must also be calibrated so that stated probabilities correspond to observed frequencies.

AI and physics are increasingly combined

The useful distinction is not simply AI versus physics. Data-driven models learn from observations and reanalysis produced through measurement, physical modelling and data assimilation. Hybrid systems can retain conservation laws, numerical solvers or physical constraints while using machine learning selectively. Ensembles represent multiple plausible futures rather than one deceptively precise trajectory.

• Compare against strong numerical baselines across seasons and regions.
• Test rare, high-impact events separately from routine weather.
• Quantify uncertainty in a form forecasters can use.
• Measure latency and resilience when observations are missing.
• Retain expert review and operational fallback systems.

AI forecasting matters when it improves trusted warnings and decisions, not merely when it wins an aggregate benchmark.

Further reading: WMO overview of artificial intelligence in weather, climate and water.