For sixty years, climate scientists have used the same basic tool: physics-based numerical models that simulate the atmosphere, ocean, and land surface by solving thousands of equations at each point on a global grid. These models are extraordinarily powerful — and extraordinarily slow. Running a century-scale climate projection can take weeks on a supercomputer. A new generation of AI climate models has changed that calculus so dramatically that researchers are now asking a different question: not just "can AI forecast the climate?" but "what can AI find in the climate that our equations never noticed?"
1,000 years in 12 hours
The first number that reframes the conversation is 1,000. That's how many years of current climate the Deep Learning Earth SYstem Model — DLESyM, built by University of Washington researchers Dale Durran and Nathaniel Cresswell-Clay — can simulate in approximately 12 hours on a single processor. A conventional supercomputer model running the equivalent simulation would require weeks and millions of CPU-hours.
DLESyM combines two neural networks: one representing the atmosphere, one the ocean. It was trained on short-range forecasts, and something remarkable emerged: the model learned how to capture seasonal and interannual variability it was never explicitly taught. It simulates tropical cyclones, the seasonal cycle of the Indian summer monsoon, and the month-to-month and year-to-year swings of mid-latitude weather patterns — matching or outperforming the leading physical models on all of these, without equations encoding the fluid dynamics that cause them.
The Allen Institute for AI's ACE2 (AI2 Climate Emulator) goes further still. Running on a single NVIDIA H100 GPU, ACE2 produces 2,200 years of climate inference in 33 hours — at roughly 1,000 times lower computational cost than conventional physical models — with forecast skill comparable to flagship physics-based seasonal forecast systems.
(ACE2 vs physics models)
one GPU in 33 hours
on a single processor
Speed this extreme is not just a convenience. It changes what kinds of science are possible. Questions that required a supercomputer allocation request are now answerable on a laptop. Ensemble runs — running the same model with slightly different initial conditions to understand uncertainty — can be done at scales previously prohibitive. Climate scientists can explore the parameter space of their models with a thoroughness that was simply out of reach six months ago.
Finding patterns the equations never showed
Speed alone would be remarkable. But a study published in Artificial Intelligence for the Earth Systems in May 2026 shows something more interesting: AI is now being used as a scientific instrument for discovery, not just a faster version of existing forecasting tools.
The research, led by Antonios Mamalakis at the University of Virginia, combined deep learning with explainable AI (XAI) to analyse one of climate science's persistent challenges: predicting seasonal precipitation months in advance. The team built models designed not just to produce accurate forecasts, but to reveal which climate signals drive those forecasts — and why.
The finding was significant. Multiple independent AI models, trained separately and applied to the same problem, arrived at the same conclusion: winter precipitation in the southern United States — Florida, Georgia, the Carolinas, Virginia — is far more predictable than in northern states, and the primary driver is El Niño and La Niña events in the tropical Pacific. The models identified a strong teleconnection between tropical sea surface temperatures and precipitation thousands of kilometres away that the existing physics-based forecast literature had not fully characterised at this resolution.
"We are entering a period where AI can become a scientific tool, not just a forecasting tool."— Antonios Mamalakis, University of Virginia, 2026
The phrase "scientific tool" is doing significant work here. A forecasting tool tells you what will happen. A scientific tool tells you why. The shift from the first to the second is the shift that happened when researchers began using AI not just to predict outputs but to interrogate the mechanisms behind those predictions — to use the model's internal representations as evidence about how the climate system actually works.
What made the University of Virginia result particularly compelling was reproducibility across models. When different AI systems trained independently arrive at the same hidden pattern, it's harder to dismiss as a statistical artefact of the training data. Multiple convergent lines of AI analysis pointing to the same teleconnection carry the same epistemic weight as multiple independent experiments confirming the same result in a laboratory.
A new generation of tools
The climate AI landscape in 2026 is no longer a single experiment — it's an ecosystem. Several systems have now passed the threshold of scientific credibility in head-to-head tests with conventional models.
- Encode atmospheric and ocean fluid dynamics as equations
- Computationally expensive — weeks on supercomputers
- Transparent mechanistic reasoning
- Well-validated over decades of use
- Struggle with spatial resolution vs. compute tradeoffs
- Learn patterns from historical climate data
- 1,000× faster at comparable or better accuracy
- Can surface hidden signals via explainability tools
- Less interpretable — "what did it learn?" is an open question
- Uncertain behaviour outside training distribution
GraphCast and FourCastNet have already demonstrated competitive 10-day weather forecasting accuracy against the European Centre for Medium-Range Weather Forecasts — long the gold standard. NASA and IBM's Prithvi-weather-climate foundation model is designed to transfer across tasks, handling everything from short-range weather to long-range seasonal projections within a single architecture.
AI climate models are trained on historical climate data. A system trained on the last century's climate patterns faces an important limitation: climate change is moving the climate system outside the range of conditions it trained on. As extreme events become more frequent and more severe, AI models trained on historical norms may underperform precisely when accurate prediction matters most — at the tail of the distribution, where lives and infrastructure are at stake.
A 2025 paper in Science Advances found that physics-based models outperform AI weather forecasts on record-breaking extreme events. Hybrid models — combining AI speed with physical constraints — are increasingly seen as the path forward.
Science at a new speed
The practical stakes of better seasonal forecasting are substantial. Drought prediction three to six months out could allow farmers to adjust crops and water use. Flood outlooks could enable infrastructure prepositions. Wildfire season projections could inform deployment of fire suppression resources before conditions deteriorate. The US Southeast — where AI models have demonstrated the highest forecast skill — is also the region most exposed to hurricane season rainfall and flash flooding from convective storms.
But the deeper implication is epistemic: AI is beginning to serve as a tool for scientific understanding, not just for prediction. The question is whether the patterns AI finds are real discoveries about how the climate system works, or sophisticated statistical correlations that break down outside the training data. Explainability methods — tools that make a neural network's internal reasoning inspectable — are the current answer to this question. When the pattern an AI identifies can be traced back to a physically interpretable mechanism, confidence in the discovery increases.
What AI climate science looks like at the end of this decade is not yet clear. What is clear is that the instrument has changed. The telescope made visible what was always there but too distant to see. The microscope revealed what was always there but too small. AI climate models are revealing what was always there in the data — patterns of connection and causality in the climate system that the equations encoded too much to find.
Buy me a coffee