NWP stands for numerical weather prediction, the method behind virtually every weather forecast you see on your phone, TV, or favorite weather app. It works by feeding current atmospheric observations into powerful computer models that simulate how the atmosphere will evolve over hours and days. These models solve physics equations governing wind, temperature, pressure, humidity, and moisture to project future weather conditions, and they now produce meaningful forecasts out to about 10 days.
How NWP Works
The atmosphere follows the fundamental laws of physics: conservation of momentum, mass, energy, and water vapor. NWP translates these laws into a set of seven mathematical equations that describe seven key variables: wind speed and direction (in three dimensions), air pressure, air density, temperature, and humidity. The core of this system is the equations governing fluid motion, which account for forces like pressure differences, Earth’s rotation, and gravity. Additional equations track how heat moves through the atmosphere and how water vapor is gained or lost through evaporation and precipitation.
To generate a forecast, these equations are solved step by step, pushing the atmospheric state forward in time. Because the equations are far too complex to solve by hand, they require supercomputers that divide the atmosphere into a three-dimensional grid and calculate changes at millions of points simultaneously.
The Three Stages of a Forecast
Every NWP forecast follows the same basic cycle, originally outlined decades ago and still used today.
The first stage is observation. Weather stations, ocean buoys, weather balloons, aircraft, radar, and satellites collect data on current atmospheric conditions around the world. Satellite data plays an especially large role. NASA’s precipitation-measuring satellites, for instance, provide microwave-based rainfall information that helps correct tropical cyclone track forecasts and feeds into short-term “nowcasting” for the next one to five hours.
The second stage is data assimilation. Raw observations are patchy and unevenly distributed, so forecasters blend them with a recent short-range forecast to create the best possible snapshot of the atmosphere right now. This step is critical because even tiny errors in the starting conditions can snowball into large forecast errors later.
The third stage is the model run itself: the computer integrates the physics equations forward in time, producing a forecast for hours, days, or even weeks ahead. After the model finishes, post-processing adjusts the raw output for local terrain and known model biases before it reaches forecasters and the public.
Supercomputers Behind the Scenes
Running a global weather model is one of the most computationally demanding tasks in science. NOAA operates two primary supercomputers, each with a capacity of 12 petaflops (12 quadrillion calculations per second), located in Virginia and Arizona. Combined with research supercomputers in several other states, NOAA’s total supercomputing capacity reaches 40 petaflops. That power allows higher-resolution models, larger sets of forecast runs, and more sophisticated physics. The European Centre for Medium-Range Weather Forecasts (ECMWF), widely considered to run the world’s best global model, maintains similarly massive computing infrastructure.
Deterministic vs. Ensemble Forecasts
A single model run from one set of starting conditions is called a deterministic forecast. It gives you one answer: rain at 3 p.m., 45°F high, winds from the northwest. The problem is that tiny uncertainties in the initial observations can lead to very different outcomes several days later.
Ensemble forecasting addresses this by running the same model dozens of times, each with slightly different starting conditions. If most of those runs agree, forecasters have high confidence. If the runs scatter widely, uncertainty is high. This approach captures something a single forecast cannot: how predictable the weather actually is in a given situation. A slow-moving high-pressure system might produce tightly clustered ensemble members, while a developing storm could send them in many directions. Ensembles also allow forecasters to assign probabilities, like a 70% chance of heavy rain or a 20% chance of freezing temperatures, rather than issuing a single yes-or-no prediction. Research consistently shows that ensemble-based probabilistic forecasts outperform single deterministic runs.
Why Forecasts Get Worse Over Time
In 1963, mathematician Edward Lorenz published a landmark paper showing that even a minuscule difference in starting conditions would eventually cause atmospheric simulations to diverge wildly. This insight, later popularized as the “butterfly effect,” established that the atmosphere is a chaotic system with an inherent predictability limit. In a follow-up paper in 1969, Lorenz argued that even with near-perfect initial conditions, rapid error growth at small scales would cap useful forecasts at roughly two weeks.
That two-week boundary has held up surprisingly well in practice, though it’s an empirical observation rather than a hard physical law. Today’s best models produce skillful forecasts out to about 10 days, with accuracy dropping steadily after that. Five-day forecasts today are roughly as accurate as three-day forecasts were 20 years ago, a testament to steady improvements in observations, computing power, and modeling techniques.
AI Models Are Changing the Field
A new generation of artificial intelligence models is challenging the dominance of traditional physics-based NWP. These systems, trained on decades of global weather data, learn the relationships between atmospheric states at different times rather than solving physics equations directly. Pangu-Weather, developed by Huawei and published in Nature, produced stronger deterministic forecast results across all tested variables compared to ECMWF’s operational system when both were evaluated against the same historical dataset. It also tracked tropical cyclones more accurately, and it did all of this more than 10,000 times faster than the traditional model.
Google’s GraphCast showed even more surprising results. With improved initial conditions, its 10-day forecast accuracy jumped by 86%, and the model demonstrated skill at predicting weather more than 33 days into the future, well beyond the traditional two-week barrier. However, AI models currently have a significant limitation: they tend to ignore the small-scale atmospheric processes that drive the predictability limit in real weather. They also struggle with rare or unprecedented events that fall outside their training data. For now, most meteorological agencies use AI models alongside traditional NWP rather than as a replacement.
Real-World Impact
NWP output drives far more than your daily forecast. NOAA uses model data to issue heat wave outlooks 8 to 14 days in advance and provides hourly forecasts, advisories, watches, and warnings as dangerous heat approaches. Its HeatRisk tool, built on NWP output, assigns a color-coded scale from zero (minor) to four (extreme) to communicate heat-related health risk up to seven days ahead.
Precipitation forecasts from NWP feed into flood warnings and help hydrologists manage reservoirs. Air quality agencies use NWP wind and temperature forecasts to predict where wildfire smoke or pollution will travel. Farmers use multi-day forecasts to time planting and harvesting. Airlines route flights around turbulence and storms using NWP data. Emergency managers rely on hurricane track forecasts, built entirely on NWP, to decide when and where to order evacuations. In each case, the value comes from the same underlying process: turning today’s observations into a physics-based projection of what the atmosphere will do next.