Weather forecasts touch nearly everything we do — from deciding whether to grab an umbrella, to routing cargo ships across the ocean, to planning flight paths and pricing insurance policies. It's easy to take that little icon on your phone for granted, but the science behind it has a surprisingly recent and dramatic history.
From Guesswork to Physics
Forecasting as we recognize it today traces back to the 1800s, but the real turning point came during World War II, when accurate predictions were critical to planning the Normandy landings. Since then, forecasts have relied on the physical laws that govern fluid dynamics — the same equations that describe how air and moisture move through the atmosphere.
There's a catch, though: the atmosphere is chaotic, and the equations describing it are nonlinear, meaning tiny errors can snowball into big ones. The arrival of modern computers in the 1950s changed the game, allowing scientists to crunch these complex calculations at speeds no human could match — a sharp departure from the old "analog" approach, where forecasters simply compared today's weather patterns to similar ones from the past.
We've Come a Long Way — But Gaps Remain
The payoff has been enormous. Five-day forecast accuracy now exceeds 90%, up from roughly 70% in the 1980s. Modern models can simulate the atmosphere at resolutions as fine as one kilometer, updated every hour. But even at that resolution, models still rely on simplified assumptions — called "schemes" — to represent processes we don't fully understand, like exactly how water vapor turns into raindrops.
One way scientists are closing that gap is a technique called data assimilation.
What Is Data Assimilation?
Here's a way to think about it. Imagine finding an old, coffee-stained photograph of a man you believe is your great-grandfather. You're not certain — but as you ask around, your mother mentions he was once photographed by a soldier; your uncle recalls the desk in the photo was his; your grandfather remembers his love of bitter coffee. Each new piece of information nudges your confidence higher, until you're all but certain the man in the photo is who you thought.
That's the essence of data assimilation, built on Bayesian statistics. In weather forecasting, scientists start with an initial "guess" — usually a set of parallel model runs, each with slightly different starting conditions to account for uncertainty in the current atmosphere. As real observations come in (say, hourly temperature readings from a weather station), the models are continually nudged toward those observations. The result is a forecast that stays closer to reality than one running unchecked.
Testing It in the Field
A research collaboration between the University of Arizona and the U.S. National Center for Atmospheric Research put this idea to the test during the 2021 Arizona monsoon season. The team used GPS antennas at 31 sites to track atmospheric moisture every five minutes, feeding that data into a high-resolution model every hour for six hours before letting the model run freely for the following 18.
Across 24 monsoon rain events, the results were encouraging: forecasts using data assimilation more closely matched observed rainfall, temperature, and atmospheric instability than forecasts without it. The findings were presented at the American Meteorological Society's 2024 conference and the European Geosciences Union's 2025 assembly, with a paper currently under second-round review at Monthly Weather Review.
What's Next: Lidar and Machine Learning
The team is now exploring how to combine GPS moisture data with readings from a MicroPulse Differential Absorption Lidar (MPD) — without the benefit of hourly weather balloon launches. It's an open mathematical question whether this is even feasible.
Meanwhile, artificial intelligence is entering the picture. With decades of observational and reanalysis data available, researchers are training machine learning models to recognize atmospheric patterns and generate forecasts of their own. A visiting PhD student is currently spending six months trying to beat a physics-based, data-assimilation-enhanced model using AI alone.
Most scientists doubt AI will fully replace physics-based models anytime soon. A more likely near-term future: AI handling the trickiest nonlinear calculations inside traditional models, making them faster and potentially more accurate — a hybrid approach that, in its own way, echoes the pattern-matching instincts of the old analog forecasters.
The bottom line: better weather forecasts don't come from a single breakthrough, but from steadily feeding models more — and better — information. Data assimilation is one of the clearest examples of that principle in action, and it may only be the beginning.