Weather computer models are mathematical simulations of the atmosphere that predict how temperature, pressure, wind, and moisture will change over the coming hours and days. They work by dividing the atmosphere into a three-dimensional grid of cells, then solving physics equations that govern fluid motion, heat transfer, and moisture at every cell, stepping forward in small time increments. The basic ideas behind this approach were conceived about a century ago, but it took the arrival of electronic computers, satellite observations, and stable numerical methods to make the approach practical.1Journal of Computational Physics. The origins of computer weather prediction and climate modeling Today, dozens of operational models run around the clock at national weather agencies worldwide, and a new generation of machine-learning models is beginning to challenge them.
How Weather Models Build a Starting Picture of the Atmosphere
A forecast is only as good as the snapshot of current conditions it starts from. Weather agencies gather millions of observations every day from radiosondes (weather balloons), surface stations, aircraft, ships, buoys, radar, and satellites. But these observations are unevenly spaced and incomplete. A radiosonde might report temperature and humidity at a single point over Denver, while a satellite captures a broad swath of cloud-top temperatures over the Pacific with no direct information about what is happening near the ground.
Data assimilation is the process that merges all these scattered observations with a model’s own short-range forecast to produce a coherent, gap-free picture of the atmosphere at a given moment. It blends what the instruments measured with what the model expected, weighting each according to its estimated reliability.2PubMed Central. Physically consistent global atmospheric data assimilation with machine learning in latent space The physical consistency of that blended state matters enormously: if the starting winds, temperatures, and pressures are not in balance with each other, the model can produce spurious storms or other artifacts as soon as the forecast begins.
Researchers are now experimenting with machine-learning approaches to data assimilation. One recent method uses a diffusion model to assimilate sparse observations directly, while another performs the blending step in a compressed “latent” representation of the atmosphere learned by a neural network, then maps the result back to the full set of atmospheric variables.3International Conference on Machine Learning. DiffDA: a diffusion model for weather-scale data assimilation These techniques are still largely experimental, but they hint at a future where the entire forecasting pipeline, from observation processing to prediction, could be handled by AI.
The Grid and What Falls Between It
Imagine a wire-frame globe with grid cells that are, say, 10 kilometers on a side. Everything smaller than that cell, including individual thunderstorm updrafts, small cumulus clouds, and turbulent eddies, cannot be directly simulated. The model only “sees” the average conditions inside each box. Designing those grids is a genuine engineering challenge: on a sphere, some grid layouts cluster resolution near the poles (where you do not need it) and spread it thin over the tropics (where you do). Researchers have explored conformal maps, icosahedral meshes, and other geometric tricks to distribute grid cells more evenly, but every design involves tradeoffs between accuracy, computational cost, and the behavior of the numerical equations on that grid.4Quarterly Journal of the Royal Meteorological Society. Horizontal grids for global weather and climate prediction models: a review
Because so many important atmospheric processes happen below the grid scale, models rely on parameterization schemes, essentially statistical recipes that approximate the collective effect of small-scale processes on the larger grid cell. Cloud microphysics is a prime example. The formation of ice crystals, the collision and coalescence of droplets, and the way a cloud interacts with radiation all occur at scales far smaller than the grid. Even high-resolution regional models that can explicitly resolve individual thunderstorms still need to parameterize the microphysics within those storms.5Advances in Atmospheric Sciences. Parameterization and Explicit Modeling of Cloud Microphysics: Approaches, Challenges, and Future Directions Getting clouds right has outsized importance because clouds strongly influence how much sunlight reaches the surface and how much heat escapes to space. One recent approach trains a machine-learning parameterization on cloud fields from a much finer-resolution reference model, then plugs that learned scheme into the coarser model to correct radiation biases.6Journal of Advances in Modeling Earth Systems. A Machine Learning Parameterization of Clouds in a Coarse‐Resolution Climate Model for Unbiased Radiation
Why Forecasts Go Wrong After About Two Weeks
The atmosphere is a chaotic system. Tiny differences in starting conditions, differences far too small to measure, amplify rapidly over time. This is the butterfly effect: in principle, the flap of a butterfly’s wings could nudge the atmosphere onto a slightly different trajectory that, days later, looks completely different from the original. Past simulations with physics-based models and theoretical work both confirm that these small perturbations grow very quickly at first, setting a hard ceiling on how far into the future any forecast can carry useful detail.7Geophysical Research Letters. Can Artificial Intelligence‐Based Weather Prediction Models Simulate the Butterfly Effect? For day-to-day weather, that ceiling is roughly two weeks. Beyond that, the specific sequence of highs and lows becomes unpredictable regardless of how powerful the computer or how dense the observations.
This does not mean longer-range outlooks are useless. Seasonal forecasts and climate projections work on a different principle: they predict the statistics of weather (will the coming winter be warmer than average?) rather than specific events (will it rain on February 12?). But for the kind of forecast you check before deciding whether to carry an umbrella, two weeks is roughly where skill fades into noise.
Ensemble Forecasting and Uncertainty
Because the atmosphere is chaotic and no observation network is perfect, a single model run from a single set of starting conditions can mislead. Ensemble forecasting addresses this by running the same model dozens of times, each with slightly different initial conditions or tweaked model assumptions, and examining how much the runs agree or disagree.8PubMed. Weather forecasting with ensemble methods If all 50 runs show rain on Thursday, you can be fairly confident. If half show rain and half show dry skies, the forecast carries real uncertainty, and a responsible forecast will say so.
The visual shorthand for ensembles is the “spaghetti plot,” a chart covered in a tangled cluster of lines, each representing one ensemble member’s predicted path for, say, a hurricane. Where the lines bunch tightly together, forecasters have high confidence in the track. Where they fan out, confidence drops. Research into how the public interprets these plots has found that spaghetti maps have a distinctive pattern of diffusion through social media during severe weather events and can help convey the range of possible outcomes more intuitively than a single “best guess” line on a map.9Computer Supported Cooperative Work (CSCW). Achieving Accuracy through Ambiguity: the Interactivity of Risk Communication in Severe Weather Events Probabilistic flood forecasts face a similar communication challenge: researchers have studied how best to visualize ensemble-based flood predictions so that emergency managers grasp the uncertainty without being overwhelmed by it.10Hydrological Processes. Visualizing probabilistic flood forecast information: expert preferences and perceptions of best practice in uncertainty communication
Major Operational Models and How They Compare
The two models most often mentioned in public weather discussions are the European Centre for Medium-Range Weather Forecasts (ECMWF) model, sometimes just called “the European model,” and the Global Forecast System (GFS) run by the U.S. National Weather Service. In head-to-head comparisons, the ECMWF model has generally demonstrated stronger skill in medium-range forecasting. During one field campaign over the Indian Ocean, for example, both models produced good one- to two-day forecasts, but at five- to fifteen-day lead times the ECMWF model was significantly better at predicting tropical rainfall patterns, large-scale circulations, and the behavior of synoptic-scale weather systems.11Journal of Geophysical Research: Atmospheres. ECMWF and GFS model forecast verification during DYNAMO: Multiscale variability in MJO initiation over the equatorial Indian Ocean
Other countries operate their own global models. The UK Met Office runs the Unified Model, Japan operates the Global Spectral Model, and agencies in Canada, France, Australia, and elsewhere maintain independent systems. These models share some physics but differ in grid design, parameterization choices, and data assimilation techniques. Having multiple independent models is valuable precisely because their errors are somewhat independent; when they agree, confidence goes up.
Newer coupled models go further by linking the atmospheric model to ocean, land-surface, and sometimes sea-ice models that exchange information during the forecast. A recent evaluation of a coupled atmosphere-land-ocean model for India’s monsoon season showed that including ocean feedback reduced systematic rainfall biases, particularly over the equatorial Indian Ocean, compared to an atmosphere-only version of the same model.12Geophysical Research Letters. An Atmosphere–Land–Ocean Coupled Global NWP System: Performance Evaluation for Potential Operational Use Coupling adds complexity and computational cost but can be essential for phenomena like tropical cyclones, where sea-surface temperature changes beneath the storm feed back into the storm’s intensity.
How Much Forecasts Have Improved
The progress over the past several decades has been striking. A long-term analysis of forecast accuracy for Melbourne, Australia, found that a three-day minimum temperature forecast in recent years had an average error of about 1.6 °C, roughly as accurate as a one-day forecast from the 1960s and 1970s. Four-day maximum temperature forecasts were more accurate than one-day forecasts from that earlier era.13Meteorological Applications. The accuracy of weather forecasts for Melbourne, Australia That pattern, each decade buying roughly a day of additional useful lead time, has been documented at agencies around the world. The gains come from every part of the system improving simultaneously: denser observations, better data assimilation, finer grid resolution, improved physics packages, and faster computers.
The Supercomputing Demands
Running a global weather model at high resolution means solving equations at millions or billions of grid points, dozens of vertical levels, and hundreds of time steps, all within a window of an hour or two so the forecast is ready before the weather it predicts actually arrives. This is among the most demanding tasks in high-performance computing. Agencies routinely operate some of the world’s largest supercomputers for this purpose. One benchmark on Japan’s TSUBAME 2.0 supercomputer, which used nearly 4,000 GPUs, achieved 145 teraflops of performance running a high-resolution mesoscale atmosphere model developed by the Japan Meteorological Agency.14Procedia Computer Science. 145 TFlops Performance on 3990 GPUs of TSUBAME 2.0 Supercomputer for an Operational Weather Prediction That was over a decade ago; current operational systems are orders of magnitude more powerful.
The sheer energy consumption of these machines is part of why the cost and carbon footprint of weather forecasting have attracted attention. Physics-based models must run their full set of equations at every grid point for every time step, which consumes enormous amounts of electricity. One analysis found that, for a use case corresponding to one year of operational forecasts, data-driven (AI) models consumed roughly 21 to 1,273 times less energy than their physics-based counterparts, depending on the model and resolution.15Weather. Energy and carbon footprint considerations for data‐driven weather forecasting models The training phase of an AI model is expensive, but it is a one-time cost that is quickly offset by the savings at inference, when the model actually generates a forecast.
Machine-Learning Models and the New Competition
The biggest recent shakeup in the field has come from AI weather models that skip the physics equations entirely. Instead of solving fluid dynamics from first principles, these models learn the mapping from current atmospheric conditions to future conditions by training on decades of historical data. Google DeepMind’s GenCast, for example, generates an ensemble of 15-day global forecasts at roughly 25-kilometer resolution and 12-hour time steps, covering more than 80 atmospheric variables, in about eight minutes. When evaluated against ECMWF’s operational ensemble system, GenCast showed greater skill on over 97 percent of the 1,320 forecast targets tested and better predicted extreme weather events, tropical cyclone tracks, and wind power output.16Nature. Probabilistic weather forecasting with machine learning
These results are genuinely impressive, but they come with caveats. AI models depend entirely on the quality of the reanalysis data they were trained on, which is itself a product of traditional physics-based models. They can struggle with rare events that are poorly represented in the training set. And when it comes to tropical cyclone intensity, as opposed to track, benchmarks have found that AI models still need additional post-processing or task-specific fine-tuning to produce skillful forecasts.17arXiv. TCBench: A Benchmark for Tropical Cyclone Track and Intensity Forecasting at the Global Scale A model that predicts where a hurricane will go but not how strong it will be when it gets there has obvious practical limitations. Still, the speed and energy efficiency advantages are dramatic enough that most major forecasting centers are actively exploring how to fold AI models into their operational pipelines, often running them alongside traditional models rather than replacing them outright.
Forecasting Weather on Other Planets
The same principles that drive Earth weather models apply to any planet with an atmosphere, and researchers have adapted terrestrial codes to simulate conditions on Mars, Titan, Venus, and Jupiter. PlanetWRF, a modified version of the widely used Weather Research and Forecasting model, demonstrated that a single numerical framework could handle atmospheres ranging from Mars’s thin carbon-dioxide envelope to Titan’s dense nitrogen atmosphere, across scales from local terrain features to global circulations.18Journal of Geophysical Research: Planets. PlanetWRF: A general purpose, local to global numerical model for planetary atmospheric and climate dynamics A newer model, planetMPAS, builds on an unstructured-grid architecture and has reproduced observed features like Mars’s seasonal carbon-dioxide cycle and polar argon enrichment, as well as Jupiter’s equatorial superrotation and banded wind patterns.19Planetary and Space Science. Unstructured grid dynamical modeling of planetary atmospheres using planetMPAS
AI is entering this arena too. A recent experiment fine-tuned Google DeepMind’s GraphCast, originally trained on Earth weather data, for Martian conditions using a synthetic climate database. With as few as ten training epochs, the adapted model began capturing Mars’s diurnal temperature cycle and could produce forecasts up to ten days out that reproduced seasonal and vertical temperature structure.20arXiv. MarsCast: Transfer Learning of AI Weather Foundation Models to Planetary Atmospheres The practical motivation is real: Mars landers and rovers are sensitive to dust storms and temperature swings, and any future human missions will need reliable short-range weather forecasts. If Earth-trained AI models can be transferred to Mars with relatively modest additional training, rapid planetary weather prediction could become routine rather than a research curiosity.