Graph Neural Network Explained: Weather Maps and Molecules

On August 6, 2026, Google DeepMind open-sourced the code and weights for its WeatherNext family and published the accompanying paper in Nature. The headline result: benchmarked against historical cyclones from 2023 through 2025, the cyclone model delivered roughly 24 extra hours of reliable warning versus leading operational systems. A three-day forecast was about as accurate as a two-day forecast used to be. Behind that is an idea worth a parent’s half hour. Here is a graph neural network explained without jargon: it is a neural network that operates on points and the connections between them, updating each point using messages from its neighbors. That one structure works for storms, molecules, road networks, and social graphs, which is why it is worth teaching your kid.

Key Takeaways

• A graph is nodes plus edges. A graph neural network updates each node by passing messages along edges, repeatedly.
• DeepMind’s original weather model, GraphCast (Lam et al., 2022), used exactly this: mesh nodes over the globe, message passing, 10-day forecasts at 0.25° in under a minute.
• The newer WeatherNext 2 and WeatherNext Cyclones use Functional Generative Networks for ensembles; DeepMind released all three variants plus a mini version under Apache 2.0 on GitHub on August 6, 2026.
• For molecules, the same framework (Gilmer et al., 2017) treats atoms as nodes and bonds as edges, and it beat handcrafted chemistry features on the QM9 benchmark.
• Kids can learn message passing with string, sticky notes, and a rule about averaging neighbors, in about twenty minutes.

What DeepMind released on August 6

The WeatherNext release covers three models: WeatherNext 2 (the general forecaster, 0.25° resolution, fine-tuned on ECMWF HRES data), WeatherNext Cyclones (also 0.25°, tuned for tropical cyclone track and intensity), and WeatherNext 2-mini, a 1° lightweight version that DeepMind says runs in a free Google Colab notebook. Code and notebooks are Apache 2.0; other materials are CC BY 4.0. The associated paper, Alet et al. (2026) in Nature, is titled “Operational Tropical Cyclone Forecasting with AI.”

The numbers, stated carefully: benchmarked on 2023 through 2025 cyclones against ECMWF-ENS for track and HWRF for intensity, the cyclone model gave forecasters on average “an extra day’s worth of predictive accuracy.” DeepMind also generates 1,000-member ensembles where operational systems typically run 50. The cyclone model runs on a coarse 28 by 28 kilometer effective input resolution (111 km for the mini variant), which is far coarser than conventional hurricane models, and still wins on track error. That is the part meteorologists find striking.

Note the honest caveat DeepMind does not hide: these are forecasts of track and intensity, not of what happens to a specific street. Storm surge, rainfall totals, and local flooding still depend on downstream models and on human forecasters at the National Hurricane Center. An extra day of lead time is worth a great deal and is not the same as certainty.

Graph neural network explained: message passing, step by step

Start with the wrong approach, because it makes the right one obvious. A convolutional network, the workhorse of image AI, assumes your data lives on a regular grid and that neighbors are the pixels immediately around you. Earth is a sphere, so a latitude-longitude grid crams nodes together at the poles and spreads them at the equator. Molecules are not grids at all; a carbon atom might have four neighbors or two, at irregular distances. Grids are the wrong shape for both problems.

A graph fixes the shape. You define nodes (a location, an atom, a person) and edges (which nodes are connected). Then you run message passing, formalized by Gilmer, Schoenholz, Riley, Vinyals, and Dahl in 2017. Each round has three steps:

1. Message. Every node computes a small vector to send along each of its edges, based on its own state and the edge’s properties (distance, bond type).
2. Aggregate. Every node sums or averages the messages arriving from its neighbors.
3. Update. Every node revises its own state using the aggregated messages, through a small neural network.

Run that once and information travels one hop. Run it sixteen times and information from a node sixteen hops away has reached you. That is how a weather graph net propagates the influence of a pressure system thousands of kilometers away without ever being told about wind. And the learned functions in steps 1 and 3 are shared across all nodes, so the model learns physics-like rules once and applies them everywhere. That weight sharing is why these models are small enough to train.

GraphCast, published by Lam and colleagues at DeepMind in 2022, built its graph as a multi-scale mesh over the globe, with short edges for local effects and long edges for distant ones, then ran message passing. The result: hundreds of weather variables predicted 10 days ahead at 0.25 degrees “in under one minute,” outperforming ECMWF’s HRES on 90 percent of 1,380 verification targets. That paper is the reason the field pivoted.

Now, an honest update on the architecture. The newer WeatherNext 2 does not simply scale up GraphCast. It uses what DeepMind calls a Functional Generative Network (Alet et al., 2025, arXiv 2506.10772), which generates ensembles “via learned model-perturbations” rather than by adding noise to inputs. DeepMind’s own repository lists GraphCast (graph neural network) and GenCast (diffusion-based) as the legacy models in the family. So the honest framing for a kid is this: graph message passing was the breakthrough that made learned weather models competitive, and current production systems have evolved past pure GraphCast while keeping the graph way of thinking about the atmosphere.

The molecule side of the story is cleaner. Treat each atom as a node carrying its element, each bond as an edge carrying its type and length, and run message passing. Gilmer et al. reported state-of-the-art results on QM9, a benchmark of molecular properties, strong enough that they suggested the field needed harder datasets. That architecture is now standard across drug discovery and materials science: DeepMind’s GNoME materials work used graph networks to screen millions of candidate crystals.

How to Teach Your Kid About Graph Neural Networks

Ages 5–8: The whisper circle with numbers

Six sticky notes on the table, each with a number, connected by lengths of string (not all connected to all). Rule: on each round, every note changes its number to the average of itself and whatever it is tied to. Do three rounds, writing the numbers each time. Your kid will watch the numbers converge, and will notice that a note tied to many others changes fast while an isolated one lags. That is aggregation, and it is the entire mechanism at a five-year-old’s level.

Ages 9–12: Draw the graph of your own week

Have your kid draw nodes for the places they go (home, school, park, grandma’s) and edges for the trips they actually take, labeling each edge with travel time. Then pose a question that requires more than one hop: “if the road to the park floods, how does that change getting to grandma’s?” They have to propagate information along edges to answer. Then show them a weather map and ask what the nodes and edges would be. Most kids get it immediately.

Ages 13+: Run the mini model

DeepMind says WeatherNext 2-mini runs in a free Colab notebook. With supervision, have your teen open the repository, read the model card, and try the notebook. Even if the code is beyond them, three concrete skills come out of it: reading a model card, understanding a resolution figure (1° versus 0.25°), and seeing that “state of the art” is a downloadable file, not a mystery. Then have them build a molecule graph on paper for caffeine and count how many message-passing rounds it takes for the nitrogen at one end to “hear from” the oxygen at the other.

The question to ask: “How many rounds of passing messages would it take for news from the far side of the graph to reach this node?”

One structure, many domains

Domain	Nodes are	Edges are	What the model predicts
Global weather	Mesh points over Earth’s surface and atmosphere levels	Neighbor and long-range mesh links	Temperature, wind, pressure fields hours to days ahead
Tropical cyclones	Same mesh, coarser (28 km effective)	Mesh links	Track, intensity, wind structure, as a 1,000-member ensemble
Molecules	Atoms, labeled with element	Chemical bonds, labeled with type and length	Energy, solubility, binding, spectroscopic properties
Crystals and materials	Atoms in a repeating unit cell	Near-neighbor distances	Stability, formation energy
Road and transit networks	Intersections or stations	Road segments or routes	Travel time, congestion propagation
Social and citation networks	People or papers	Follows or citations	Community membership, influence spread

The table is the lesson. When a kid sees one mechanism solve six unrelated problems, they have learned something about how science actually generalizes, which is a rarer insight than any single fact about storms.

What to actually do at home

Give your kid the word “graph” in its real sense

Most kids hear “graph” and think bar chart. The mathematical sense, points and connections, is one of the most useful concepts in all of computing and it is entirely accessible to a nine-year-old. Two evenings of sticky notes and string buys a concept they will reuse in chemistry, geography, and any programming course.

Use hurricane season as the teaching window

When a storm is in the news, pull up the National Hurricane Center’s forecast cone with your kid and talk about what the cone means: it is the spread of an ensemble, not the size of the storm. This is the single most misread graphic in American weather coverage, and DeepMind’s 1,000-member ensembles make the concept more relevant, not less.

Separate “faster” from “better”

GraphCast’s headline was speed: minutes on one machine versus hours on a supercomputer. That is real and it is not the same as accuracy. Have your kid articulate which one they care about for a school-cancellation decision. Precision about what a number measures is the actual skill.

Connect it to what they already like

A kid who plays with molecular model kits is already building graphs. A kid who maps subway systems is building graphs. A kid who draws social diagrams of their class is building graphs. Naming the structure they already play with raises the ceiling on what they can learn next. Our pieces on the open-sourced WeatherNext 2 cyclone model and how AI is changing weather forecasting go further.

What not to do

Do not tell your kid that AI has replaced meteorologists. The National Hurricane Center still issues the warnings, human forecasters still adjust the guidance, and the models are inputs. Overstating this is how kids end up distrusting the warning that matters.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can define nodes and edges and explain what “one round of message passing” does.
• Month 2 red flags: Your kid says “AI predicts the weather now” without being able to name what it predicts (fields, not street-level flooding).
• Month 3 self-check: During the next storm in the news, can your family read the forecast cone correctly as an ensemble spread? If yes, the concept landed.

Frequently Asked Questions

Is a graph neural network the same as a neural network?

It is a family of neural networks with a specific structure: the connections in the data determine which units talk to which. A standard fully connected network ignores structure; a convolutional network assumes a grid; a graph network reads the structure you give it. Same math underneath, different assumption about shape.

Why is this better for weather than a normal image model?

Because Earth is a sphere and the atmosphere has multiple levels, so a flat pixel grid distorts distances badly near the poles. A mesh graph lets you place nodes evenly and connect both nearby and distant points explicitly, which matches how atmospheric influence actually travels.

Can my kid actually run one of these models?

The mini cyclone model is designed to run in a free Colab notebook, so technically yes with supervision. Realistically, a middle schooler will get more from the paper-and-string version plus reading the model card. Save the code for a motivated high schooler.

Does this mean AI weather forecasts are always better?

No. The published comparisons are strong on cyclone track and on many global variables, and weaker or unproven for other things, including local rainfall extremes. Operational forecasting blends multiple models. Treat a single model’s win on a benchmark as evidence, not a verdict.

What careers use this?

Weather and climate modeling, computational chemistry, drug discovery, materials science, logistics, and fraud detection all use graph learning. The common prerequisite is linear algebra plus programming, and the differentiator is knowing a domain deeply enough to define the right graph.

Is “open source” here the real thing?

Mostly. Code and notebooks are Apache 2.0, model weights are downloadable, and the paper is published. Running the full models wants serious hardware (an H100-class GPU or a TPU). Reproducing the training from scratch is out of reach for almost everyone, which is the standing limit of “open” in frontier AI.

---

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. Google DeepMind. (2026, August 6). “AI model achieves breakthrough in forecasting cyclones.” https://deepmind.google/blog/weathernext-ai-model-achieves-breakthrough-in-forecasting-cyclones/
2. Alet, F., et al. (2026). “Operational Tropical Cyclone Forecasting with AI.” Nature. https://www.nature.com/articles/s41586-026-10953-2
3. Lam, R., Sanchez-Gonzalez, A., Willson, M., et al. (2022). “GraphCast: Learning skillful medium-range global weather forecasting.” arXiv 2212.12794. https://arxiv.org/abs/2212.12794
4. Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., & Dahl, G. E. (2017). “Neural Message Passing for Quantum Chemistry.” arXiv 1704.01212. https://arxiv.org/abs/1704.01212
5. Alet, F., et al. (2025). “Skillful joint probabilistic weather forecasting from marginals.” arXiv 2506.10772. https://arxiv.org/abs/2506.10772
6. Google DeepMind. “WeatherNext model repository.” GitHub. https://github.com/google-deepmind/weathernext
7. NOAA National Hurricane Center. “Understanding the forecast cone.” https://www.nhc.noaa.gov/aboutcone.shtml