AI Learns to Predict Earth's Weather Patterns
AI predicts weather patterns with unprecedented accuracy by analyzing massive datasets, enhancing climate models and reducing future uncertainties for a.
By Sci Herald Editorial Team ·
Based on research by Thomas Wilder, Ranjini Swaminathan, David Storkey, Till Kuhlbrodt, first posted to arXiv in November 2026 — view the original study
TL;DR: A new AI system can predict the churning, small-scale ocean currents that drive global weather patterns with remarkable accuracy. This advance could lead to more reliable long-range weather forecasts and sharper climate change projections for communities worldwide.
Key Facts
- The AI model predicts the energy of ocean eddies—swirling currents that move heat around the planet—with up to 95% accuracy.
- It was trained on a massive, high-resolution simulation of a simplified Atlantic Ocean, learning from four distinct oceanic regions.
- The research tackles a core problem in weather forecasting: standard computer models are too coarse to see these small but powerful currents, so they must guess their effects.
- The system shows promising ability to generalize, performing well in some unseen ocean regions if their conditions resemble its training data.
- This work is a foundational step toward integrating this AI into the UK's official Earth System Model used for national climate projections.
The Tiny Ocean Swirls That Shape Our Weather
Imagine trying to predict the path of a hurricane without understanding the behavior of every small wave and current in the ocean that fuels it. For climate scientists, that's the daily challenge. Global weather and climate models divide the planet into a grid. But the supercomputers that run them aren't powerful enough to calculate what's happening in every single square. The grids are too big, leaving countless small-scale processes—especially in the ocean—unseen.
These unseen processes are not minor details. They are the engines of our climate system. The most important are ocean eddies: massive, swirling vortices of water, tens to hundreds of kilometers across, that act like conveyor belts moving heat from the equator toward the poles. Their combined effect is enormous, but since they are smaller than the grid boxes in most global models, their influence has to be approximated—a process known as parameterization. Getting this guesswork right is one of the biggest sources of uncertainty in predicting our future climate.
Now, a team of researchers from the University of Reading and the UK Met Office is using artificial intelligence to tackle this very problem. In a recent paper, Thomas Wilder, Ranjini Swaminathan, David Storkey, and Till Kuhlbrodt describe an AI model that learns to predict the hidden energy of these ocean eddies with record precision. Their work, published on the preprint server arXiv, represents a significant step toward building smarter, more trustworthy climate models.
How Does AI Predict Weather Patterns?
AI predicts weather patterns by learning the hidden rules of the atmosphere and ocean from vast amounts of simulation data, effectively filling in the gaps that supercomputers alone cannot calculate.
The researchers used a type of AI called a convolutional neural network—the same technology that helps your phone recognize faces in photos. But instead of recognizing smiles, this AI was trained to recognize the complex fluid dynamics of the ocean.
Their training ground was a sophisticated video game-like simulation of the ocean called the DIabatic Neverworld Ocean (DINO). It's a simplified, idealized version of the Atlantic Ocean, designed to be a perfect testbed for this kind of research. They ran this high-resolution simulation for the equivalent of 60 years to generate a planet's worth of data on how ocean currents, temperatures, and energies behave.
The map above shows the simulated "Neverworld" ocean, with colors representing ocean depth. The labeled boxes (JET, TROP, etc.) are the distinct regions—one resembling a fast-moving current, others calm gyres—where the AI was trained. The N-JET region in the north was completely held back from training to test the AI's ability to generalize.
The team then created the AI's homework. They took the detailed simulation data and blurred it, effectively lowering its resolution to mimic what a standard climate model can actually see. The AI's job was to look at these blurred, low-resolution snapshots of ocean conditions—things like water velocity and pressure—and predict the true, high-resolution energy of the eddies that were missing from the picture. It's like showing someone a pixelated image of a forest and asking them to correctly guess the number and type of trees in every blurry patch.
Teaching an AI to Generalize
A critical question for any AI tool is whether it can work in situations it wasn't specifically trained for. Can a model trained on Atlantic Ocean dynamics predict what happens in the Pacific? This ability is called generalization, and it's vital for building a truly useful global climate AI.
The team rigorously tested this by training several different AI models. Some were trained on data from all four oceanic regions in their simulation. Others were trained on just three regions, with one region completely held out to serve as a final exam. They discovered a key insight: The AI could successfully generalize to an unseen region only if the physical conditions there were similar to those in its training data.
For example, an AI trained on data from a fast, turbulent jet stream could make good predictions about another fast, turbulent region it had never seen. But if that same AI was asked to make predictions about a calm, slow-moving ocean gyre, its performance dropped significantly. The models encountered something called distributional shift—the test data was too different from the homework it had studied.
"ML parameterization models generalize to unseen regions when their training data share similar feature and target distributions with those regions," Thomas Wilder and colleagues write in the paper.
This finding is crucial. It means that for an AI weather prediction system to be reliable everywhere on Earth, its training dataset must be very diverse, capturing the full spectrum of the planet's climates and geographies. This research helps scientists understand how to build that comprehensive dataset. This approach of learning from data is similar to how other AI systems are built, like those that predict business futures by analyzing market patterns.
The Performance: Stunning Accuracy with a Catch
The results from the best-performing AI model were impressive. When predicting the overall average energy of ocean eddies, the model achieved a near-perfect score of 95% accuracy (an R² score of ~0.95). For predicting changes in that energy over time (anomalies), it scored between 80% and 90% accuracy.
However, the study also honestly addressed the model's limitations. The AI sometimes struggled to accurately predict the energy of the very smallest scales of motion it was tasked with modeling. Furthermore, the researchers found that the choice of which ocean features to show the AI (a process called feature selection) involved trade-offs. Using a certain set of data could improve predictions in one region but slightly worsen them in another. There is no one-size-fits-all solution, underscoring the complexity of the Earth's systems.
*This diagram from the research paper shows a snapshot of the ocean data the AI processes. More accurate parameterization of ocean eddies means climate models can make better predictions about long-term processes: how ocean currents will shift, how much heat the ocean will absorb, and how quickly sea levels will rise.
This leads to reduced uncertainty in climate projections. For policymakers, this means having more confident answers to critical questions: How severe will droughts in this region be in 2050? How much more frequent will extreme rainfall events become? For the public, it means communities can make better-informed decisions about building sea walls, managing water resources, and protecting agriculture—all based on more trustworthy climate forecasts. The goal of creating AI that can operate reliably in complex, physical environments is a theme across different fields, from climate science to robotics, as seen in research into how AI learns to grab and move objects like humans.
The Bigger Picture
This study is part of a massive, ongoing shift in climate science. Researchers are increasingly turning to machine learning to solve problems that have traditionally stumped purely physics-based approaches. The world is simply too complex to model perfectly with equations alone. By feeding AIs with immense amounts of data, scientists hope to uncover patterns and relationships that we haven't even thought to write equations for.
The work by Wilder and team is a specific contribution to a larger goal: building a machine-learned component for the UK's Earth System Model (UKESM). This is the model the UK government uses to inform its climate policy and commitments to international agreements like the Paris Accord. Improving its core components with AI could have a tangible effect on national and global climate strategy.
The field is moving rapidly. Other teams are using AI to tune existing physics formulas or to discover entirely new equations hidden in the data. The common thread is a hybrid approach—using AI not to replace our hard-won physical knowledge of the climate, but to augment it, creating models that are both physically consistent and data-informed. This mirrors a broader trend in AI development towards creating more capable and efficient models, much like the comparisons drawn between different on-device AI systems for developers.
The open challenge, which this paper directly addresses, is ensuring these powerful AI tools don't become "black boxes" that work without explanation. By testing their generalization and probing their failures, researchers are building the trust required to put them into the operational systems that forecast our future.
Frequently Asked Questions
Can AI actually predict the weather better than traditional methods?
For short-term forecasts of the next few days, traditional physics-based models are still extremely effective. The promise of AI is in improving long-range forecasts and climate projections, where it can help model very complex small-scale processes, like ocean eddies, that traditional methods have to simplify and approximate.
Does this mean human meteorologists will be replaced by AI?
No. AI will be a powerful tool that assists meteorologists and climate scientists. It can handle the immense number crunching required to simulate countless small-scale interactions. Human experts are still essential for interpreting the results, understanding the limitations of the models, and communicating the forecasts and their uncertainties to the public.
How soon will we see this technology in use?
This particular research is foundational. The next step is to take this AI, which was trained "offline" on pre-recorded data, and actually plug it into a live, running climate model to see how it performs over time. This integrated testing is a major hurdle that will take several more years of research and development before it could be considered for use in official government climate models.
What is the biggest hurdle for using AI in climate science?
The biggest hurdle is trust and reliability. Climate scientists must be absolutely certain that an AI component will behave in physically plausible ways under all conditions, especially the extreme conditions of a future warming climate that it may not have been explicitly trained on. Research like this, which rigorously tests an AI's ability to generalize, is key to building that trust.
Source: Read the original study