There’s a quiet miracle that happens every morning before most people wake up. Supercomputers in a handful of countries finish churning through quadrillions of calculations, satellites beam down images of clouds from orbit, weather balloons transmit data from the upper atmosphere, and all of it gets distilled into the little icon on your phone that tells you whether to bring a jacket.
We complain about the forecast when it’s wrong, but we rarely stop to appreciate how often it’s right. A five-day forecast today is as accurate as a one-day forecast was thirty years ago. That improvement didn’t happen by accident. It’s the result of a global infrastructure of observation, computation, and modeling that represents one of humanity’s greatest scientific achievements.
Here’s how we actually know what the weather is going to do, and the technologies that make modern forecasting possible.
The Foundation: Observation Everywhere, All at Once
A weather model is only as good as the data you feed into it. The atmosphere is a fluid, chaotic and interconnected, and small errors in knowing its current state grow rapidly into large errors in predicting its future. This is the famous butterfly effect, and it’s why meteorologists are obsessive about gathering as much observational data as possible, from as many places, at as many levels of the atmosphere.
The backbone of this observational network is the surface station. Tens of thousands of automated weather stations dot the planet, at airports, on buoys, on remote mountaintops, measuring temperature, pressure, humidity, wind speed, wind direction, and precipitation. These stations report continuously, feeding a constant stream of ground truth into the forecasting system.
But the surface is only the bottom of the atmosphere. What’s happening higher up matters just as much, because weather systems are three-dimensional. This is where radiosondes come in. Twice a day, at the same time, hundreds of weather stations around the world release weather balloons carrying instrument packages that rise through the atmosphere, transmitting temperature, humidity, and pressure data as they ascend, tracked by GPS to measure wind speed and direction at every level. These balloons reach the stratosphere before they burst, and the data they send back is irreplaceable. Satellites can see a lot, but nothing replaces an instrument physically moving through the air column.
Aircraft provide another layer of upper-air data. Thousands of commercial flights are equipped with sensors that report temperature and wind conditions along their routes. The data is transmitted automatically, and it fills in gaps over oceans and continents alike. Every plane descending into a major airport is also sending back a vertical profile of the atmosphere, a free radiosonde in all but name.
The ocean surface, where hurricanes are born, is monitored by an array of buoys and drifters. Moored buoys provide continuous data from fixed locations. Drifting buoys, deployed by ships and aircraft, move with the currents and report via satellite. Argo floats go deeper, descending to two thousand meters and then rising, measuring temperature and salinity through the water column. The ocean stores an enormous amount of heat, and understanding that heat content is essential to forecasting everything from tropical cyclones to seasonal climate patterns.
Satellites: The View from Orbit
If surface stations and weather balloons are the skeleton of the observing system, satellites are the skin. They provide the continuous, global view that nothing else can.
Geostationary satellites sit in a fixed position relative to the ground, parked about thirty-six thousand kilometers above the equator. They rotate with the Earth, staring at the same hemisphere continuously. The GOES series operated by the United States, Meteosat operated by Europe, Himawari operated by Japan, and others provide overlapping coverage of the entire planet. These satellites take images every few minutes, in visible light, infrared, and water vapor channels. The loops they produce, showing clouds swirling and storms developing in real time, are the most familiar satellite product, but they also provide quantitative data on cloud-top temperatures, atmospheric moisture, and wind vectors derived from tracking cloud motion.
Polar-orbiting satellites fly much lower, a few hundred kilometers up, and they sweep from pole to pole as the Earth rotates beneath them, covering the entire planet twice a day. The joint NASA-NOAA Suomi NPP and NOAA-20 satellites carry instruments like ATMS and CrIS, which measure temperature and humidity at different levels of the atmosphere by detecting microwave and infrared emissions from the air itself. This is called sounding, and it’s essentially a vertical atmospheric profile from space. For areas without weather balloon coverage, which is most of the ocean and much of the developing world, polar satellite soundings are the only regular data on what the atmosphere is doing aloft.
Newer satellites carry lightning mappers that track electrical discharges in storms, providing data on storm intensity and rapid intensification. Scatterometers use radar to measure wind speed and direction at the ocean surface, critical for tropical cyclone forecasting. Altimeters measure sea surface height, which reveals the heat content below. The constellation of Earth-observing satellites is one of the most underappreciated pieces of infrastructure in modern civilization.
Radar: The Ground-Based Eye on Storms
Satellites see the big picture. Radar sees the details. Weather radar sends out pulses of radio waves that bounce off precipitation and return to the antenna. The time it takes for the signal to return tells you how far away the precipitation is. The strength of the returned signal tells you how intense it is. The shift in frequency, via the Doppler effect, tells you how fast it’s moving toward or away from the radar.
The United States operates a network of over 150 NEXRAD Doppler radar sites, providing near-complete coverage of the country. Other nations operate similar networks. Dual-polarization radar, which sends and receives pulses in both horizontal and vertical orientations, is a recent advance that has improved precipitation estimation and made it possible to distinguish between rain, snow, hail, and debris lofted by tornadoes. This technology has made tornado warnings more accurate and increased lead times.
Phased-array radar, currently in development and testing, replaces the traditional rotating dish with a flat panel of small antennas that can be steered electronically. This allows a volume scan of a storm to be completed in under a minute, compared to several minutes for conventional radar. In a rapidly evolving severe weather situation, those minutes matter. Phased-array radar provides a higher-resolution view of storm evolution and could further improve tornado warning performance.
Supercomputers: Where the Data Becomes a Forecast
All the observations in the world are useless without a way to project them forward in time. This is the job of numerical weather prediction, and it requires some of the most powerful computers on the planet.
A weather model divides the atmosphere into a three-dimensional grid of millions of cells, each described by a set of variables: temperature, pressure, humidity, wind, and so on. The model applies the equations of fluid dynamics, thermodynamics, and radiative transfer to each cell, calculating how it will change in a series of small time steps. This process is repeated over and over, stepping forward through time to produce a forecast.
The Global Forecast System, or GFS, run by NOAA, covers the entire globe with a grid spacing of about thirteen kilometers. The European Centre for Medium-Range Weather Forecasts, or ECMWF, runs a model that’s widely considered the global gold standard, with a grid spacing down to nine kilometers. Both are run on supercomputers that occupy entire buildings and consume megawatts of electricity.
Because the atmosphere is chaotic, no single model run is sufficient. Forecasters use ensemble forecasting, running the same model multiple times with slightly different initial conditions to represent the uncertainty in the observations. If all fifty members of an ensemble show a hurricane making landfall in the same place, confidence is high. If they scatter across a hundred-mile stretch of coastline, confidence is lower. Ensembles provide not just a forecast but a measure of its uncertainty, which is arguably more valuable than the forecast itself.
High-resolution regional models nest inside the global models, zooming in on specific areas with much finer grid spacing. The High-Resolution Rapid Refresh, or HRRR, covers the continental United States with a three-kilometer grid, updating every hour. This model captures small-scale features that the global models blur out, including individual thunderstorm cells and terrain effects. It’s the model that tells forecasters whether a line of storms will hold together long enough to reach your town.
Artificial Intelligence: The New Frontier
The newest tool in forecasting is not a physical instrument but a software one. Artificial intelligence and machine learning are being applied to weather prediction in ways that are advancing rapidly.
Traditional numerical weather prediction solves the equations of physics from first principles. AI models take a different approach. They train on decades of historical weather data, learning the patterns and relationships without being explicitly programmed with the physics. Once trained, they can produce a forecast in seconds on a modest computer, compared to hours on a supercomputer for traditional models.
Google DeepMind’s GraphCast model has demonstrated forecast skill competitive with the ECMWF on certain metrics, producing ten-day global forecasts in under a minute. Huawei’s Pangu-Weather and NVIDIA’s FourCastNet are similar efforts. These models don’t replace traditional forecasting, at least not yet, but they are already being used alongside physics-based models, and the gap is closing.
The most promising near-term application is post-processing. Machine learning models can take the output of traditional weather models and correct their systematic biases, learning that, for example, a certain model tends to under-predict nighttime temperatures in certain geographic regions. AI is also being used for nowcasting, very short-term forecasts of the next hour or two, where it can analyze radar and satellite imagery faster than traditional methods and produce updates in near real-time.
The challenge is that AI models learn from past data, and climate change means the future atmosphere won’t behave exactly like the past. A model trained on thirty years of data may struggle with conditions it hasn’t seen before, which are becoming more common. The field is evolving fast, and the likely future is a hybrid approach, physics-based models and machine learning models working together, each covering the weaknesses of the other.
The Human Element: Why Forecasters Still Matter
With all this technology, it’s reasonable to wonder whether human forecasters are still necessary. The answer is yes, emphatically, particularly for high-impact events.
Numerical models and AI forecasts provide the raw material. Forecasters interpret it. They know the local terrain, the local microclimates, the systematic biases of different models. They know that a certain model tends to over-intensify storms in a certain region, or that another model struggles with fog in a particular valley. They synthesize the model output, the satellite imagery, the radar, and their own experience into a coherent story.
During severe weather events, this synthesis is life-saving. A forecaster can see a tornadic signature on radar, correlate it with the atmospheric environment, and issue a warning with context that a purely automated system would struggle to provide. They can communicate uncertainty in ways that help emergency managers make decisions. The human is not a weak link in the forecasting chain but the final integrator of an almost impossibly complex flow of information.
The Bottom Line
Modern weather forecasting is a worldwide machine that never sleeps. Satellites stare down from orbit. Balloons rise through the atmosphere. Radar dishes scan the skies. Supercomputers run the physics. AI models learn the patterns. And human forecasters, standing in front of screens full of data, turn all of it into words that help people decide whether to evacuate a coast, whether to plant a crop, whether to carry an umbrella.
The accuracy of a five-day forecast is now comparable to what a one-day forecast was a generation ago. Lives have been saved by improvements in hurricane track forecasting and tornado warning lead time. The technology is not perfect, and it never will be, because the atmosphere is fundamentally chaotic beyond a certain horizon. But the quiet miracle continues, every morning, as the data flows in and the forecasts go out. We are, collectively, better at predicting the future than we have ever been. The sky tells its story, and now, more than ever, we know how to listen.
