The AI Now Generating Your Family's Weather Forecast

The weather app on your phone doesn’t tell you which computer generated your forecast. That matters, because the answer has changed.

For most of the past 50 years, weather forecasts came from numerical weather prediction (NWP) models: massive simulations of Earth’s atmosphere, running the equations of fluid dynamics and thermodynamics on supercomputers. They’re physics-based, painstakingly validated against decades of observation data, and expensive to run. The European Centre for Medium-Range Weather Forecasts (ECMWF) in Reading, UK, operates one of the world’s most powerful dedicated supercomputers — just to run weather models.

In 2023, that started changing. Google DeepMind published results for GraphCast, a machine learning model trained on 40 years of historical weather data. It outperformed the ECMWF’s flagship model on 90% of weather variables and produced 10-day global forecasts in under one minute on a single Google TPU chip. ECMWF’s model needs hours of supercomputer time to do the same thing.

This is a real scientific shift — not a marketing claim — and it has practical consequences for hurricane prediction, climate modeling, and the careers your kid might end up pursuing.

Why Parents Should Care About This

Most parents think of weather forecasting as a mature, stable field. It’s on television. It’s been on television since television existed. The weatherman hasn’t changed much.

What most parents don’t see is that weather forecasting is actually one of the most computationally intensive, data-hungry problems in all of science. Atmospheric science sits at the exact intersection of physics, mathematics, data science, and planetary-scale systems thinking. It was already demanding. AI has now made it one of the fastest-moving fields in applied machine learning.

For a kid interested in climate, natural disasters, or math — this is a field that is being genuinely reinvented. The next generation of atmospheric scientists will be expected to understand both the physics of the atmosphere and the machine learning models being applied to it.

What the Research and Data Show

The performance of AI weather models was definitively established in a landmark 2023 paper published in Science by Google DeepMind’s Remi Lam and colleagues. GraphCast was evaluated against the ECMWF’s high-resolution forecast (HRES) system — widely considered the gold standard — on 1,380 verification targets. GraphCast outperformed HRES on 90.3% of those targets.

The comparison was rigorous: 39 years of ERA5 reanalysis data (the most comprehensive historical weather dataset available) for training, with verification on held-out data from 2018 forward. This wasn’t an easy comparison — HRES is a decades-refined, physics-based model. GraphCast beat it in a fraction of the compute time.

Pangu-Weather, developed by Huawei Cloud and published in Nature in 2023 by Bi and colleagues, achieved comparable results — outperforming the ECMWF model on most standard atmospheric variables and running 10,000× faster. Critically, Pangu-Weather demonstrated superior performance on cyclone track forecasting over a 3-month evaluation period.

NVIDIA’s Earth-2 platform takes a different approach: it combines ML models with high-resolution physics-based simulation to target extreme weather events specifically. NVIDIA announced in 2023 that Earth-2’s CorrDiff diffusion model could generate 25km-resolution weather predictions from lower-resolution inputs at 1,000× speedup over traditional downscaling approaches.

Model	Developer	Published	Approach	Key result vs. ECMWF HRES
GraphCast	Google DeepMind	2023 (Science)	Graph neural network trained on ERA5 reanalysis data	Outperforms on 90% of variables; runs in <1 min on a single chip
Pangu-Weather	Huawei Cloud	2023 (Nature)	3D Earth system transformer	Outperforms on most variables; 10,000× faster; better hurricane tracks
FourCastNet	NVIDIA	2022 (preprint)	Fourier neural operator	Comparable accuracy; designed for ensemble runs at high speed
AIFS	ECMWF	2024	ML model developed in-house	ECMWF’s own AI model, now running operationally alongside HRES
GenCast	Google DeepMind	2024 (Nature)	Probabilistic diffusion model	Outperforms ensemble forecasts at 15-day range

The Physics-Based vs. Machine Learning Distinction — What Kids Should Understand

This is the conceptually important part, and it’s worth explaining clearly enough that parents can explain it to their kids.

How traditional weather models work

Numerical weather prediction starts with physics. The atmosphere is a fluid — a very large, compressible, rotating fluid. The equations that describe how fluids move (Navier-Stokes equations) can be applied to the atmosphere if you know the current state (temperature, pressure, humidity, wind speed) at enough points.

In practice, you divide the atmosphere into a three-dimensional grid of cells — maybe 10km × 10km horizontally, with layers at different altitudes. You measure the current state of each cell using weather stations, radiosondes (weather balloons), aircraft sensors, and satellites. Then you step the equations forward in time — each step is maybe 15 minutes of simulated time — and repeat thousands of times to project days ahead.

The problem: it’s enormously expensive computationally, and small errors in the initial state (measurement uncertainty) grow over time. This is the fundamental reason weather forecasts degrade past about 10–14 days — a limitation called the “predictability limit” established by meteorologist Edward Lorenz in the 1960s.

How AI weather models work differently

GraphCast and its cousins don’t simulate physics directly. Instead, they learn statistical relationships from historical data. ECMWF’s ERA5 dataset contains atmospheric state information for the entire globe, from 1940 to present, at 6-hour intervals. GraphCast was trained to predict: given the atmospheric state at time T and T-6 hours, what will the state be at T+6 hours?

The model learns these patterns through billions of training examples — implicitly learning many physical relationships without being told the equations explicitly. Then at inference time (generating a new forecast), it applies what it learned to current atmospheric data, iterating forward in 6-hour steps.

This is fundamentally different from simulation. It’s pattern-matching on an enormous corpus of historical atmospheric behavior. Which is also why experts are still debating what happens near the boundaries of that historical distribution — extreme climate scenarios that haven’t appeared in the training data may not be predicted well.

The honest answer is: these models are stunning for medium-range weather forecasting (3–10 days). How well they handle the unusual events that matter most for climate change is still being studied. That’s exactly the kind of unsettled science that creates interesting careers.

What This Means for Hurricane Prediction

Hurricane track forecasting improved dramatically over the past 40 years using traditional NWP. AI models are now showing they can match or exceed that — but rapid intensification (the dangerous phenomenon where a hurricane suddenly strengthens by 35+ knots in 24 hours) remains difficult for both traditional and AI models.

NOAA’s Hurricane Research Division is actively evaluating how AI models can be integrated into operational hurricane forecasting. The 2023 Atlantic hurricane season was one of the first where AI model output was routinely available to NWS forecasters alongside traditional model guidance.

For a kid who grows up curious about hurricanes, climate disasters, or atmospheric science, the career path ahead involves genuinely unsolved scientific problems — not just optimizing existing methods.

What Parents Should Do

1. Watch weather forecasting as a teaching tool

Weather is one of the few scientific phenomena kids encounter every day. When a forecast is right or wrong, that’s an opportunity. “Why did they say it would rain but it didn’t? What data did they have? What did they get wrong?” This teaches the idea that models are approximations and understanding their failure modes is as important as understanding their predictions — a key insight in AI generally.

2. Introduce the concept of training data

GraphCast was trained on 40 years of observed weather. That’s the historical record the model learned from. Ask your kid: what would happen if the historical record was incomplete? What if the climate changes so much that the future doesn’t look like the past? This is a real open question in atmospheric AI research, and it’s accessible to a curious 11-year-old.

3. Explore free atmospheric data tools

NOAA provides enormous amounts of free atmospheric data — including the output of their operational weather models — through the NOAA Open Data Dissemination program. NASA’s Earthdata portal provides satellite-based atmospheric data. These aren’t toys; they’re the actual data scientists use. A high school student with Python skills can do real atmospheric data analysis with them. Our article on future-proofing kids’ careers with AI skills covers the programming foundation that makes this possible.

4. Connect climate interest to data science

A kid who cares about climate change and wants to “do something about it” has a direct path through atmospheric data science. Climate models are a type of numerical weather prediction, extended to decadal timescales. The people who build, validate, and improve them are applied mathematicians, physicists, and increasingly, machine learning engineers. All of it is a skills story, not just a values story.

5. Follow ECMWF, NOAA, and NASA as sources

These institutions publish genuinely accessible explainers about how their models work, what they’re improving, and where the challenges are. ECMWF’s blog in particular is well-written and technically substantive. Following real scientific institutions teaches kids what scientific communication looks like, which is different from what most science YouTube looks like.

6. Watch for local university atmospheric science programs

Many land-grant universities have atmospheric science or meteorology departments that offer summer programs for high school students. The National Center for Atmospheric Research (NCAR) in Boulder, Colorado, and NOAA’s various laboratories run internship and mentorship programs. These exist and are competitive but accessible.

What to Watch Over the Next 3 Years

AI models entering operational forecasting — ECMWF began running its own AI-based model (AIFS) operationally alongside HRES in 2024. NOAA has announced plans to integrate ML model guidance into National Weather Service operational products. Within 3 years, the weather forecast on your phone will almost certainly be partially or fully AI-generated in the background, even if the interface looks the same.

Sub-seasonal to seasonal forecasting — The 2–4 week range (beyond typical 10-day forecasts but shorter than climate projections) is called sub-seasonal to seasonal (S2S) forecasting and has historically been the hardest range. AI models are showing early promise here. Improvements in this range would have enormous value for agriculture, energy demand planning, and disaster preparedness.

High-resolution extreme event prediction — Current AI weather models generate global forecasts at roughly 0.25-degree resolution (about 28km at the equator). Useful, but not fine enough for county-level severe weather prediction. The next challenge is downscaling — using AI to generate high-resolution predictions from coarser model output. NVIDIA’s CorrDiff is an early example; more will follow.

Climate model emulation — The same ML techniques used for weather forecasting are being applied to climate projections — running models that approximate the output of slow, expensive physics-based climate models in a fraction of the time. This would allow climate scientists to explore far more scenarios than they currently can, improving the confidence and resolution of climate projections.

If atmospheric science has historically felt like a closed field for kids who aren’t already on a meteorology path, the AI transformation is reopening it. Machine learning engineers from software backgrounds are now working at ECMWF, NOAA, and Google DeepMind alongside physicists who spent careers on numerical weather prediction. The field needs both.

Frequently Asked Questions

Is the weather forecast on my phone now using AI?

Most major consumer weather apps (Weather.com, Weather Underground, Apple Weather) blend traditional NWP model output with AI-powered post-processing and localization. As of 2024–2025, AI models like GraphCast are providing additional guidance that human forecasters use alongside traditional model output. Full AI-primary forecasting in consumer products is beginning to roll out.

Can AI predict weather better than traditional models?

For medium-range forecasts (3–10 days), yes — benchmarks published in Science and Nature show AI models outperforming ECMWF’s flagship system on most standard variables. For rare extreme events and at very short ranges (0–2 hours, called “nowcasting”), traditional and specialized physics-based models often still have advantages. It’s not either-or; most operational forecasting centers are integrating both.

What subjects should a kid study to work in atmospheric AI?

Mathematics (particularly differential equations and statistics), physics, and computer science form the core. Many atmospheric science programs have added data science and ML coursework to their curriculum. A kid who’s strong in math and curious about climate is an excellent candidate for this field, even without any prior meteorology background.

How is climate modeling different from weather forecasting?

Weather forecasting predicts the specific state of the atmosphere at a specific time and place. Climate modeling projects statistical properties of the atmosphere — average temperatures, precipitation patterns, extreme event frequencies — over decades. They use some of the same physics-based tools, but the questions and validation methods are quite different. AI is now being applied to both.

Are AI weather models reliable for hurricane forecasting?

AI models have shown competitive hurricane track forecasting results, with Pangu-Weather demonstrating superior cyclone tracking in head-to-head evaluation. Rapid intensification prediction — the hardest part — remains challenging for AI and traditional models alike. NOAA and other agencies are actively testing AI model hurricane guidance operationally.

Where do weather models get their starting data?

From a combination of weather stations, ocean buoys, weather balloons, aircraft automatic reports, radar, and most importantly — satellites. NOAA, EUMETSAT, and other agencies operate geostationary and polar-orbiting weather satellites that provide continuous coverage of the global atmosphere. Getting accurate initial conditions (called “analysis”) is itself a complex data assimilation problem.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Lam, R., Sanchez-Gonzalez, A., Willson, M., et al. (2023). “Learning skillful medium-range global weather forecasting.” Science, 382(6677), pp. 1416–1421. https://doi.org/10.1126/science.adi2336

2. Bi, K., Xie, L., Zhang, H., et al. (2023). “Accurate medium-range global weather forecasting with 3D neural networks.” Nature, 619, pp. 533–538. https://doi.org/10.1038/s41586-023-06185-3

3. Price, I., Sanchez-Gonzalez, A., Alet, F., et al. (2024). “GenCast: Diffusion-based ensemble forecasting for medium-range weather.” Nature, 637, pp. 84–90. https://doi.org/10.1038/s41586-024-08252-9

4. Pathak, J., Subramanian, S., Harrington, P., et al. (2022). “FourCastNet: A global data-driven high-resolution weather model using adaptive Fourier neural operators.” arXiv preprint. https://arxiv.org/abs/2202.11214

5. ECMWF. (2024). AIFS — ECMWF’s machine learning forecast model. European Centre for Medium-Range Weather Forecasts. https://www.ecmwf.int/en/forecasts/documentation-and-support/aifs-ecmwf-artificial-intelligence-forecasting-system

6. NVIDIA Corporation. (2023). Earth-2: An AI Supercomputer for Climate and Weather Simulation. NVIDIA Technical Blog. https://developer.nvidia.com/blog/nvidia-earth-2-cloud-platform-uses-generative-ai-for-10x-more-detailed-regional-climate-modeling/

7. National Oceanic and Atmospheric Administration. (2024). Hurricane Research Division: AI and Machine Learning in Hurricane Forecasting. NOAA. https://www.aoml.noaa.gov/hrd/

8. Lorenz, E. N. (1969). “Predictability of a flow which possesses many scales of motion.” Tellus, 21(3), pp. 289–307. https://doi.org/10.3402/tellusa.v21i3.10086