Table of Contents
Weather Model Engineer Career: Inside WeatherNext 2
Google open-sourced a cyclone model that gained a day of lead time. The weather model engineer career behind it has a real path, with honest pay data.
A three-day hurricane forecast from Google DeepMind’s WeatherNext Cyclones model is about as accurate as a two-day forecast was before it. One extra day. For a family on the Gulf Coast or in the Caribbean, one day is the difference between an orderly evacuation and a scramble. The Nature paper published August 6, 2026 calls that gain roughly a decade’s worth of historical forecasting progress, delivered in one model. Then Google open-sourced it. The weather model engineer career is the job of building things like that, and unusually for a frontier AI field, a 14-year-old can run a version of the actual model this weekend.
Key Takeaways
- WeatherNext Cyclones predicts tropical cyclone track, intensity, and wind structure with about a day or more of additional lead time compared with prior systems, published in Nature on August 6, 2026 and described as equivalent to a decade of historical forecasting progress.
- It was trained on nearly 20 terabytes of global atmospheric data plus the IBTrACS historical cyclone database covering roughly 5,000 past storms.
- Google DeepMind open-sourced the code and weights on GitHub under an Apache 2.0 licence in three variants: WeatherNext Cyclones, WeatherNext 2, and WeatherNext 2-mini, which runs on a single TPU in a public Colab notebook.
- The work was co-developed with the U.S. National Hurricane Center, the Cooperative Institute for Research in the Atmosphere, and the UK Met Office, and was used during the 2025 hurricane season to help forecast Hurricane Melissa’s rapid intensification.
- Honest pay picture (BLS, May 2025): atmospheric scientists including meteorologists earn a median of $99,070 with only 3% growth and 10,700 jobs, while data scientists earn $120,230 with 35% growth and computer and information research scientists earn $140,300 with 22% growth. The AI side of weather sits in the second and third rows.
How an AI weather model actually works
Traditional weather forecasting solves physics. You divide the atmosphere into a grid, write down the equations of fluid motion and thermodynamics, and step them forward in time on a supercomputer. It works, it’s expensive, and its accuracy is limited by grid resolution and by how well you know the starting conditions.
A machine-learning weather model does something different. It learns a mapping from “the state of the atmosphere now” to “the state of the atmosphere in six hours” by training on decades of historical weather data. It never writes down Navier-Stokes. It learns the patterns those equations produce.
The counterintuitive result is that this is both faster and, for some variables, more accurate. Faster because running a trained network is cheap compared to integrating a physics model. More accurate in places because the network can learn systematic biases in observations that a physics model treats as noise.
For cyclones specifically, the team combined general atmospheric training data, nearly 20 terabytes of it, with IBTrACS, the historical cyclone database covering roughly 5,000 storms. That second dataset is why the model is good at storms specifically rather than at weather in general.
The honest limits: a learned model can only be as good as its training distribution, so genuinely unprecedented events are the hardest case, which is exactly when forecasts matter most. And the collaboration with the National Hurricane Center, CIRA, and the Met Office exists because a model is not an operational forecast until a human agency with legal responsibility says so. Our explainer on the open-sourced WeatherNext 2 cyclone model covers the release in more detail.
What a weather model engineer does day to day
Data engineering. Atmospheric reanalysis datasets are enormous and awkward. Getting 20 terabytes into a form a training pipeline can consume, at the right resolution, on the right grid, is genuinely most of the work.
Model architecture. Deciding how to represent the atmosphere to a neural network. Graph neural networks are common because the atmosphere is naturally a mesh of interacting points.
Training and evaluation. Running training jobs, then scoring against held-out years. The evaluation is harder than the training: you have to decide what “better” means for a forecast, and different users care about different errors.
Verification against operations. Comparing model output to what the National Hurricane Center actually issued and to what happened. This is where partnership with an operational agency becomes essential.
Open-source maintenance. After a release under Apache 2.0, someone answers issues, fixes the Colab notebook, and helps a meteorological service in another country get the model running.
Communication. Explaining to a forecaster, an emergency manager, or a journalist what “a day of additional lead time” means and does not mean.
The weather model engineer career path: school to a modelling role
| Stage | What to take or earn | Cost and time | Entry roles | What that role does daily |
|---|---|---|---|---|
| High school | Physics, calculus if available, statistics, earth science; Python; run the WeatherNext 2-mini Colab notebook | Free | Volunteer weather observer (CoCoRaHS), school weather station lead | Recording observations, plotting data |
| Bachelor’s | Atmospheric science, physics, applied mathematics, or computer science. Must include fluid dynamics or partial differential equations plus programming and statistics | 4 years | Junior forecaster, research assistant, data engineer | Running models, cleaning data, making figures |
| Certification | No licence required. Useful: a national meteorological society certification for broadcast or consulting work, cloud platform certificates for the modelling side | Weeks to months | Operational forecaster, cloud data engineer | Shift forecasting, pipeline maintenance |
| Master’s (common) | MS in atmospheric science, climate science, or machine learning applied to earth systems | 1–2 years | Research scientist, model developer | Building and evaluating models |
| PhD (the research tier) | Atmospheric physics, applied mathematics, or ML for weather and climate | 4–5 years, usually funded | Research scientist at DeepMind, NOAA, ECMWF, a national weather service, or a university | Designing architectures, publishing, deciding the research direction |
| Alternative entry | Computer science degree plus self-taught atmospheric physics | 4 years | ML engineer on a weather team | Training infrastructure, evaluation tooling |
| Day to day | Wrangle data, train models, verify against operations, maintain open source, explain results to forecasters | — | — | — |
The alternative-entry row is the one most families underestimate. Teams building these models need people who can handle 20-terabyte training pipelines, and that skill is scarcer than atmospheric physics knowledge. A computer science graduate who reads a meteorology textbook is currently more hireable on these teams than a meteorologist who never learned to code.
Pay and demand, said plainly
| Category | Median pay (May 2025) | Projected 2025–2035 | Jobs (2025) | Typical education |
|---|---|---|---|---|
| Computer and information research scientists | $140,300 | +22% | ~38,600 | Master’s |
| Software developers | $135,980 | +10% | ~1,700,000 | Bachelor’s |
| Data scientists | $120,230 | +35% | ~275,600 | Bachelor’s |
| Atmospheric scientists, incl. meteorologists | $99,070 | +3% | ~10,700 | Bachelor’s |
Read that table honestly. “Meteorologist” is a small occupation, about 10,700 positions, growing 3%. The AI weather work is being done inside research-scientist and data-scientist job codes that pay more and grow faster. Anyone quoting a salary for “weather model engineer” specifically is estimating from postings at a handful of employers: Google DeepMind, NVIDIA, Microsoft Research, NOAA, ECMWF, national weather services, and a small number of private forecasting companies.
The realistic advice for a kid: get the computational degree, add the atmospheric physics, and accept that the number of frontier positions is in the hundreds globally. The consolation is that the same skills move to climate modelling, energy forecasting, insurance catastrophe modelling, and agricultural technology, all of which are much larger employers.
What a 10–15-year-old can do this year, free
Run WeatherNext 2-mini in Colab
This is the unusual one. Google published a public Colab notebook that runs the mini variant on a single TPU, free. A motivated 14-year-old can produce a forecast from an actual frontier weather model in an afternoon. Almost no other frontier AI system is this open.
Join CoCoRaHS or run a school weather station
CoCoRaHS is a volunteer precipitation observation network used by the National Weather Service. Your kid’s daily rainfall measurement goes into a real dataset that forecasters use. Our rain gauge and weather station project has the build, and the discipline of daily observation is the actual professional habit.
Verify a forecast for 30 days
Write down the official high temperature forecast for tomorrow, every day, then record what happened. After a month, calculate the average error. Your kid will discover that forecasts are better than people think and worse than people assume, and they’ll have done forecast verification, which is a real job function.
Read the IBTrACS database
The historical cyclone database is public. Have your kid look up a storm that affected somewhere they care about and trace its track and intensity. Then ask what a one-day-earlier forecast would have changed for that storm.
What not to do
Don’t let them conclude that AI replaced physics models. The operational forecast still comes from a human agency that blends multiple models, including physics-based ones, and the WeatherNext work was co-developed with the National Hurricane Center and the Met Office rather than around them. And don’t skip statistics: the entire field runs on quantifying uncertainty, and a model developer who can’t explain a probabilistic forecast can’t do the job. Our piece on how AI is changing weather prediction for families is a good read-together.
What to Watch For Over the Next 3 Months
- Week 4: Your kid has run the Colab notebook and produced any output at all, and has 20 days of forecast-verification records.
- Month 2 red flags: They stop the verification log. They want the model to predict something dramatic rather than to be measured. They treat the model’s output as truth rather than as a forecast with an error bar.
- Month 3 self-check: Ask them what the average error in their 30-day verification was, and whether the forecast was better for temperature or for rain. If they can answer, they’ve done the work that underlies every claim in that Nature paper.
Frequently Asked Questions
Is this a real career path or a handful of jobs at Google?
Both, uncomfortably. The frontier positions number in the hundreds globally across DeepMind, NVIDIA, Microsoft Research, ECMWF, NOAA, and national weather services. But the skills, large-scale data engineering, graph neural networks, forecast verification, transfer directly to energy load forecasting, insurance catastrophe modelling, and agricultural forecasting, which employ far more people.
Does my kid need to study meteorology?
Not necessarily. Many people on these teams came from computer science or applied mathematics and learned atmospheric physics afterwards. What is non-negotiable is fluid dynamics or partial differential equations at some point, plus statistics, plus real programming ability.
Does open-sourcing the model mean anyone can forecast hurricanes now?
Anyone can run the model. Issuing an official forecast is a different thing: national weather services hold that responsibility legally, and they blend multiple models with forecaster judgment. The open release matters most for meteorological agencies in countries that could not afford their own modelling effort.
Will AI models put forecasters out of work?
BLS projects 3% growth for atmospheric scientists, which is neither a boom nor a collapse, and that projection already accounts for expected technology change. The tasks changing are model running and routine forecast production. The tasks that aren’t: deciding what to communicate, to whom, when a storm threatens a specific city.
What’s the single best thing to do first?
Open the Colab notebook. It takes an hour, it’s free, and it converts “AI does weather” from a headline into something your kid has personally executed. Everything else in this field is easier to motivate after that.
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
- Google DeepMind. (2026). “WeatherNext: AI model achieves breakthrough in forecasting cyclones.” August 6, 2026, with the accompanying Nature paper. https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/
- Google DeepMind. WeatherNext open-source repository (Apache 2.0). https://github.com/google-deepmind/weathernext
- Google. (2026). “Google AI updates, August 2026.” https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
- Bureau of Labor Statistics. (2026). “Atmospheric Scientists, Including Meteorologists.” Occupational Outlook Handbook, May 2025 data. https://www.bls.gov/ooh/life-physical-and-social-science/atmospheric-scientists-including-meteorologists.htm
- Bureau of Labor Statistics. (2026). “Data Scientists.” May 2025 data. https://www.bls.gov/ooh/math/data-scientists.htm
- Bureau of Labor Statistics. (2026). “Computer and Information Research Scientists.” May 2025 data. https://www.bls.gov/ooh/computer-and-information-technology/computer-and-information-research-scientists.htm
- Bureau of Labor Statistics. (2026). “Software Developers, Quality Assurance Analysts, and Testers.” May 2025 data. https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm
- National Oceanic and Atmospheric Administration. International Best Track Archive for Climate Stewardship (IBTrACS). https://www.ncei.noaa.gov/products/international-best-track-archive
